Beruflich Dokumente
Kultur Dokumente
Context Appraisal
Defended by
Diana Di Lorenza Arellano Tavara
Thesis Advisors
Dr. Francisco J. Perales
Dr. Javier Varona
2012
2
i
ii
Dr. Francisco Jose Perales Lopez.
HACEN CONSTAR:
Que la memoria titulada Visualization of Affect in Faces based on Context Appraisal ha sido
realizada por Diana Di Lorenza Arellano Tavara bajo nuestra direccion en el Departamento
de Matematicas e Informatica de la Universitat de les Illes Balears y constituye la tesis
para optar al Grado de Doctor en Informatica.
iii
Dr. Francisco Jose Perales Lopez
Director de la tesis
iv
To all who contribute to make this Thesis a reality.
v
vi
Acknowledgments
Muchsimas gracias
vii
viii
Abstract
Virtual Characters are more than avatars capable of expressing emotions and interact
with the users. Virtual Characters should be seen as a very reliable representation of a
human being, capable of expressing all the possible affective traits after the appraisal and
evaluation of what is happening around and inside them. They should feel and express
what they are feeling, they should convince you they are real.
To achieve this level of believability several researchers have proposed different com-
putational and affective models, as well as graphical techniques to simulate expressions,
gestures, behavior or voice. All this state of art has provided us with sufficient data and
information to see what else needs to be done.
As a result, we propose a contextual and affective framework that allows the generation
of the context that surrounds the character as well as the simulation of its psychologi-
cal characteristics like preferences, standards, personality, or admiration for other agents.
Moreover, the framework proposes novel and implementation independent techniques for
the visualization of emotions and mood.
Through experimentation we come up with a set of head-position/eye-gaze configu-
rations that are perceived as certain personality traits, we validate the generation of ex-
pressions for moods, and assessed the feasibility of the context generation through movie
scenes, which translated into our system, triggered the same emotions and elicit the same
facial expressions as in the movie.
This research is a step forward in the creation of more believable virtual characters, by
pointing out other elements that should be considered when creating characters that can
be used in affective HCI applications, storytelling, or virtual worlds for entertainment (e.g.
videogames) or for therapies (e.g. in therapies with autistic children).
ix
x
Resumen
Hablar de personajes virtuales implica hablar de mucho mas que avatares capaces de expre-
sar emociones e interactuar con los usuarios. Los personajes virtuales deberan ser vistos
como una representacion fidedigna de los seres humanos, capaces de expresar un amplio
rango de rasgos afectivos despues de haber analizado y evaluado que ocurre fuera y dentro
de ellos. Deben sentir y expresar lo que sienten de tal forma que logren convencer que son
reales.
Para alcanzar este nivel de credibilidad gran cantidad de investigadores han propuesto
diferentes modelos afectivos y computacionales, as como tecnicas en graficos para simular
expresiones, gestos, comportamientos y voz. Todo este trabajo previo nos ha permitido
obtener suficientes datos para analizar que mas se puede hacer en esta area.
Como resultado, proponemos una metodologa que permite la generacion automatica del
contexto que rodea al personaje, as como la simulacion de sus caractersticas psicologicas
como preferencias, estandares, personalidad, o admiracion por otros agentes. Mas aun, se
presentan novedosos algoritmos independientes de la implementacion para la visualizacion
de emociones y humor.
Mediante experimentos y test que miden el grado de percepcion en los usuarios asoci-
amos un conjunto de configuraciones orientacion de la cabeza/direccion de la mirada a
rasgos de personalidad, y validamos el metodo para generar expresiones de humor. Tambien
evaluamos la fiabilidad de la generacion de contexto usando escenas de pelculas, obteniendo
el mismo set de emociones y expresiones faciales que en dichas pelculas.
Finalmente, cabe destacar que este trabajo de investigacion es un paso hacia adelante
en la creacion de personajes mas crebles, ya que indica que elementos deberan tomarse
en cuenta al momento de crear personajes virtuales que puedan ser usados en aplicaciones
Interaccion persona-ordenador, cuentacuentos, o mundos virtuales destinados al entreten-
imiento (videojuegos) o fines medicos (terapias con ninos autistas).
xi
Palabras Claves: Personajes virtuales, Representacion de contexto, Expresiones faciales,
Psicologa.
xii
Contents
1 Introduction 1
1.1 The nature of the problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Motivation - The Domain of Interest . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Aims of Research . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.4 Research Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.5 Significance and Potential Applications . . . . . . . . . . . . . . . . . . . . . 6
1.6 Thesis outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2 Psychological Theories 9
2.1 Psychological Theories of Emotion . . . . . . . . . . . . . . . . . . . . . . . 9
2.1.1 Categorical Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.1.2 Dimensional Models . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.1.3 Appraisal Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.2 Psychological Theories of Personality . . . . . . . . . . . . . . . . . . . . . . 17
2.2.1 Eysenck Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.2.2 Five Factor Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.2.3 Circumplex Structures . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.3 Psychological Theories of Mood . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.3.1 Ekman . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.3.2 Pleasure-Arousal-Dominance Space . . . . . . . . . . . . . . . . . . . 22
2.3.3 UWIST Mood Adjective Checklist . . . . . . . . . . . . . . . . . . . 23
2.3.4 Positive and Negative Affect . . . . . . . . . . . . . . . . . . . . . . 23
2.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
xiii
CONTENTS
4 Framework Overview 47
4.1 System Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
4.2 Semantic Layer: Context Representation . . . . . . . . . . . . . . . . . . . . 50
4.2.1 Context - Inner world of the character . . . . . . . . . . . . . . . . . 50
4.2.2 Context - Outer world of the character . . . . . . . . . . . . . . . . . 51
4.3 Affective Layer: Affective Model . . . . . . . . . . . . . . . . . . . . . . . . 52
4.3.1 Emotions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
4.3.2 Mood . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
4.3.3 Personality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
4.4 Visualization Layer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
5 Context Representation 55
5.1 Context - An Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
xiv
CONTENTS
xv
CONTENTS
8 Evaluation 147
8.1 Objectives of Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
8.2 Experiment: Visualization of Emotions . . . . . . . . . . . . . . . . . . . . . 148
8.2.1 Hypothesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
8.2.2 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
8.2.3 Experimental Study . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
8.2.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
8.3 Experiment: Visualization of Moods . . . . . . . . . . . . . . . . . . . . . . 155
8.3.1 Hypothesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
8.3.2 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
8.3.3 Experimental Study . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
8.3.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
xvi
CONTENTS
9 Conclusions 187
9.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
9.2 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
9.3 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
9.4 Publications and contributions . . . . . . . . . . . . . . . . . . . . . . . . . 193
9.4.1 Journals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
9.4.2 Proceedings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
9.4.3 Workshops . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
9.4.4 Research placements . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
Bibliografy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
xvii
CONTENTS
xviii
Chapter 1
Introduction
If we knew what it was we were doing, it would not be called research, would it?
Albert Einstein.
Since the creation of the first virtual character, a lot of research has been done to
provide them with realism, believability, and empathy.
Nowadays, thinking of virtual characters means thinking of virtual worlds and
videogames as L.A. Noire, Call of Duty, The SimsTM , among many others. These charac-
ters are endowed with great realism and believability thanks to advanced technologies in
Computer Graphics and Artificial Intelligence.
Nevertheless, these techniques might be computationally expensive and therefore, not
very suitable to implement in real time interactive virtual worlds. From the AI point of
view, they should interact with the user or other inhabitants, and have affective reactions
to a series of events. From the physical point of view, they are human or animal repre-
sentations with gestures, voices, and facial expressions that show their affective states in
different instants of time.
These characters or avatars (virtual representations of a person) are not only limited
to videogames or entertainment; they can be used almost in any interactive application as
virtual presenters, educational tutors, instructors, or social networks avatars.
Nevertheless, to engage an audience, the characters must be believable, specially when
it comes to their affective responses. In this regard, the field of Affective Computing has
made great advances to give characters affective characteristics. Affective computing is a
1
CHAPTER 1. INTRODUCTION
term coined by Rosalind Picard [123] that relates to, arises from, or deliberately influences
emotions. Affective computing includes implementing emotions, and therefore aids to test
different theories of emotions. It also includes giving a computer the ability to recognize
and express emotions.
Given that our goal is to develop a framework for creation of believable characters
capable of a wide range of facial expressions, elicited as consequence of their emotional
reactions to the events in a changing virtual world, is one of the key fields in this research.
In this chapter, the nature of the problem, motivation, aims of research, and research
methodology are introduced. Also, the significance of the chosen approaches and possible
applications are identified. Finally, the thesis outline is summarized at the end of the
chapter.
One of the main problems to face in Affective Computing is the lack of consensus in answers
to questions like: what are emotions?, or what is personality?. As Picard stated, there
is open debate about these topics, and evidence lacks on all sides of the debates. Like her,
we have based our work on relevant theories, and how they have been used for creating
believable characters.
But what is believability? Paiva et al. [121] observed that believability is one of the most
debated properties of synthetic characters and the goal of researchers working on this area
for many years now. The term was introduced by Bates team [11] relating to characters
that give the illusion of life, facilitating the users suspension of disbelief. Believability
has been intensively explored in literature, and it is still the Holy Grail of the synthetic
characters research area. Why are synthetic characters not believable? Is it too hard?
Moreover, the question what makes a character behave in a believable way? arises
from the appraisal of situations and events that the characters experience. Therefore, a
precise and complete description of what surrounds the characters and how they perceive
it is necessary to make them react accordingly, and show the feelings they are experiencing.
Finally, how to evaluate the expression of affect in the character in a believable way can
only be measured empirically and subjectively with a significant sample of subjects. To
this respect the obtained results will be valid as long as the hypothesis formulated proves
to be true, or false, depending on the case that wants to be evaluated.
2
1.2. MOTIVATION - THE DOMAIN OF INTEREST
The research presented in this thesis is motivated by the idea of having a platform where
one can create interactive virtual characters and situations automatically and straightfor-
wardly. To achieve this, integration is the key. Thus integration of semantics, affective
computing and computer graphics is the basis of a system that allows the representation
of what happens with the characters, how they feel, and how they express those feelings.
The first domain of interest is semantics, which help us to define context. The main
reason for using context is stated by Kaiser and Wehrle [78] in the following paragraph:
The current, concrete meaning of a facial expression can only be interpreted within the
whole temporal and situational context. In everyday interactions, we know the context
and we can use all information that is available to interpret the facial expression of another
person. Therefore, if we generate the context we can generate accurate facial expressions
according to it.
The second domain of interest is driven by the contexts affective approach, which
constitutes one of the novelties of our work. By defining the affective traits of the character
as part of the context, a more accurate affective state of the character, and thus more
accurate facial expressions will be achieved.
The third domain of interest is focused on the characters facial expressions. Having a
character whose facial cues evoke the facial behavior of human beings can be of great help
to enrich the transmission of the contexts affective message.
Therefore, we propose a three-layered model, where the first two layers, a semantic
layer and an affective layer, are the ones that deal with the context. The semantic layer
defines the context (at a characters internal and external level) and produces an affective
output that is interpreted by the affective layer. The affective layer provides the psycho-
logical background to evaluate the emotions, mood and personality of the character, and
transform them into a representation for facial expressions. The third layer deals with the
visualization of emotions and moods, which constitutes one of the novelties of the work.
So far emotions were the main affective traits to be shown through facial expressions. Nev-
ertheless, mood is also an important affective trait that can be manifested in the face of a
person.
Another affective trait that has been poorly researched when it comes to its facial
expression is personality. Personality, by definition, is stable; but as Linda Edelstein said,
put a character in extraordinary circumstances, and certain traits come to the forefront
3
CHAPTER 1. INTRODUCTION
while others recede [40]. Nevertheless, people tend to show the same traits when placed
in similar situations: A highly competitive man will likely show ambition in the office,
or playing Monopoly with his family. The perception of personality based on observation
has long been a subject of research in behavioral psychology. It is just until recently that
this research has focused on facial actions. For this reason, as part of the third domain of
interest and as a novel research, we explore some facial cues to express personality.
A model that defines the concepts that are part of the context of the character
(its outer (environment) and inner (psychology traits) world).
A methodology that allows the user to define and infer knowledge about the
context, and to create new scenarios in a simpler way.
Psychology-based rules to produce emotional responses in the characters.
An affective model that uses psychology theories of affect to elicit new affective
states in the character based on its felt emotions, personality and mood.
A mathematical representation of affective traits to be computationally
tractable.
4
1.4. RESEARCH METHODOLOGY
Context Representation
Affective Model
1. How can we obtain facial expressions for universal and intermediate emotions?
2. Which facial cues should be considered when expressing mood and personality?
3. Do the physical characteristics of the face influence the recognition of an affective
trait?
By doing the corresponding research to answer the former questions we will be capable
of choosing the best techniques and methods to implement and validate the Computational
Affective Model proposed in this thesis.
5
CHAPTER 1. INTRODUCTION
6
1.6. THESIS OUTLINE
characters, which we will use as a reference for our research. The discussion focuses on the
techniques, theories, and results obtained by each previous computational model, and how
they serve as the basis for our own computational model.
Chapter 3 provides an overview of the framework for the computational affective mod-
ule. The discussion aims to give readers a general vision of the whole model and guide
them into the subsequent chapters in this thesis. In particular, the system architecture
will be presented to shortly introduce the different modules to be developed.
Chapter 4 introduces the semantic model used to represent context. There it is analyzed
the motivation for using context and the previous works that have researched on semantic
techniques. We also present our requirements and methods for context representation, how
to use this model to create stories, definition of characters, their environment and their
emotional responses.
Chapter 5 explains the affective model used for the computation of the characters emo-
tional states from personality, mood and emotions values. It takes as input the characters
emotional responses generated by the semantic model, and produces a new mood using
the characters personality traits and previous mood. The chosen representation for these
affective traits is based on psychological theories and affective models presented in Chapter
2.
Chapter 6 describes the visualization module which is used to generate facial expressions
for the affective state of the character. For visualization of mood, which constitutes one of
the novelties of this work, an in-depth explanation if offered so it can be replicated by future
researchs. Finally, our contribution to the research on facial expressions for personality is
described.
Chapter 7 reports the evaluation of the Computational Affective Model. It exposes the
obtained results which validates not only the effectivity of the computational model, but
the correct visualization and perception of the elicited facial expressions.
Chapter 8 summarizes our work, provides an outlook to its potentials and implications,
analyses the limitations of the taken approaches and gives some directions for future work.
7
CHAPTER 1. INTRODUCTION
8
Chapter 2
Psychological Theories
Psychology has been one of the base research fields in Affective Computing, because it
provides the affective models and theories to be used.
As our main goal is to create believable and affective virtual characters, in the follow-
ing we outline a selection of psychological theories focused on representation of affective
components as emotions, mood, and personality. This selection has been guided by the
importance and contribution of these works to the generation of virtual characters.
The study of emotions is a challenging area, since emotions can be analyzed from different
perspectives. This has originated a number of theories and models that intend to explain
what they are, as well as how and why they are appraised and elicited. This section
attempts to classify and overview some of the theories that have been used in the com-
putational field, thus we have: Categorical models, Dimensional models, and Appraisal
models.
9
CHAPTER 2. PSYCHOLOGICAL THEORIES
Darwin
Charles Darwins work is so relevant because he made major contributions to the study
of facial expressions in a way that had not been done before. In his book The Expression
of Emotions in Man and Animals [33], he stated that facial expressions and involuntary
movements are based on three principles: serviceable associated habits (certain movements
are done even when they are not necessary, e.g. scratch ones head when thinking or when
being confused); antithesis (perform movements of a directly opposite nature when having
a directly opposite state of mind, e.g. move the arms to wave away a person even if that
person is not close enough); and direct action of the nervous system (certain expressions
are influenced by physiological reactions).
By means of multiple observations in several countries using as subjects infants, people
with dementia, Duchenes studies, art works, and people from different cultures and races,
he studied how people behave when experiencing different affective states. For example
when suffering; in anxiousness, pity or despair; when feeling happiness or devotion; and
so on. For example, he observed that laugh or smile were expressions for the state high
spirits in a deaf and blind person, a normal person and idiots (medical term).
Darwin grouped emotions in categories according to shared characteristics and move-
ments, focusing primarily on the face. He grounded the idea that facial expressions of emo-
tion are universal and gestures are culture-specific conventions, also seen in other species,
which evolved serving once particular functions (e.g. baring teeth in anger to prepare for
attack), becoming useful when communicating these emotions to others.
Ekman
Paul Ekman, inspired by Darwins approach [46], studied the universality of emotional
expressions and developed a methodology to describe these expressions based on muscular
movements, the Facial Action Coding System - FACS [48].
From his experiments, Ekman confirmed Darwins theory of universality [45], claiming
that the same emotion might be elicited by different circumstances, but its expression could
10
2.1. PSYCHOLOGICAL THEORIES OF EMOTION
Plutchik
Plutchik also accounts for emotions that are either combinations of two or three basic
emotions, or one basic emotion experienced at a greater or a milder intensity. He called this
combination dyad, e.g. joy and acceptance produce love. Although Plutchik was aware that
some combinations might never occur at all, he stated that his model covered all aspects
of emotional life. However, the model is questionable when trying to classify concepts as
anticipation or surprise.
11
CHAPTER 2. PSYCHOLOGICAL THEORIES
Whissell
Cynthia Whissell provided a list of emotional terms compiled in her Dictionary of Affect in
Language [153]. It includes approximately 4000 English words with affective connotations,
where each word is described along the dimension of Activation (or Arousal) and along the
dimension of Evaluation (or Pleasantness).
Whissells work is used for measuring emotion, and though its lower reliability in the
Activation dimension, it has been proved to work better when applied to passages or lists
because it allows the evaluation of the affective tone of the entire passage or list.
In practice, the Dictionary of Affect can be applied to both short-term and long-term
responses (mood description, personality description, reaction to immediate situations, and
analysis of texts or diaries). As words are rated along a two-dimensional space, Whissell
observed that the classification of words as emotional is related to their distance to the
origin.
Russell
Another theory based in a two-dimensional bipolar space is the one proposed by Rus-
sell [135]. Based on previous experiments performed by colleagues, Russell also found
that there are three properties of the cognitive representation of affect: the pleasantness-
unpleasantness and arousal-sleep dimensions; the bipolarity of dimensions that describe
affect; and that any affect word could be defined as a combination of the pleasure and
arousal components.
As a result, a two-dimensional space was evaluated with the horizontal dimension cor-
responding to pleasure-displeasure, and the vertical corresponding to arousal-sleep. Russell
also observed the lack of need for a third or more dimensions, because having extra dimen-
sions would only account for a tiny proportion of the variance and are limited to subsamples
of emotion words. Figure 2.2 shows the eight affect concepts in a circular order.
12
2.1. PSYCHOLOGICAL THEORIES OF EMOTION
In recent works, Russell [133] proposed a framework that weds bipolar dimensions to
discrete categories. He presented a set of concepts that intended to re-arrange knowledge
from previous models to make a consensus among them. The first set of concepts are
technical terms which define various emotion-related events: Core affect, Affective qual-
ity, Attributed affect, Affect regulation and Object. The second set of concepts bridge the
gap between the technical terms and folk concepts, leading to a more familiar manner of
speaking: Mood, Empathy, Displeasure motive, Prototype, Emotional episode, Prototypical
emotional episode, Emotional meta-experience and Emotion regulation.
Cochrane
Thomas Cochrane [29] proposed an eight-dimensional model to map the conceptual space
of emotions as faithfully and efficiently as possible. We included this model due to its
potential to be used in computational applications given that it offers a useful tool for
researchers, regardless of the theory of emotions that they hold. It applies equally to what-
ever component of emotion (appraisals, emotion language, subjective feeling, physiological
changes, expressive behaviors, action tendencies or regulation strategies), integrating dif-
ferent approaches by capturing the meaning of the emotion at an abstract level.
The eight proposed dimensions are: Valence (attracted-repulsed), Personal Strength
(powerful-weak), Freedom (free-constrained), Probability (certain-uncertain), Intentional fo-
cus (generalized-focused), Temporal flow (future directed-current-past directed), Temporal
Duration (enduring-sudden), and Social connection (connected-disconnected).
13
CHAPTER 2. PSYCHOLOGICAL THEORIES
After mapping emotional terms in his model, Cochrane proved that even terms re-
lated to emotional subclasses can be differentiated, and none of these differences would be
captured by the traditional valence or arousal dimensions.
According to the appraisal theory of emotions, the emotional responses results from a
dynamic evaluation (appraisal) of needs, beliefs, goals, concerns, environmental demands
that might occur consciously or unconsciously. Therefore, this type of theories has become
one of the most active, and attractive approach in the domain of emotional psychology.
OCC Model
One of the most used model of emotions in the computational field is the one proposed by
Ortony, Clore and Collins [120], known as the OCC Model. This model is of a cognitive
nature, and intends to explain peoples perception of the world and how it causes them to
experience emotions.
For Ortony et al., emotions cannot be arranged in a low-dimensional space; rather they
should be organized in groups. They found representative clusters identified by eliciting
conditions, under which emotions are triggered. Also, inside each emotion group, each
emotion type is seen as a family of closely related emotions.
The assumption of the model is that there are three major aspects of the world, upon
persons can focus: events, agents, or objects, which elicit different types of emotions. When
one focuses on events is because of their consequences, when one focuses on agents is because
of their actions, and when one focuses in objects is because of their aspect or properties.
The structure of the OCC model is shown in Figure 2.3, where individual groups of emotion
types are enclosed in boxes, with the groups name in the bottom part of the box.
The intensity of emotions for each group is given by a number of variables that de-
pend on the appraisal of the event, agents or objects. For instance, FORTUNES-OF-
OTHERS, there are four variables that affect the intensity of its emotions: desirability-for-
self, desirability-for-other, deservingness, and liking.
Regarding the OCC model, Bartneck [9] [10], reflected on the missing features (extensive
amount of knowledge to categorize the affective response, history function to keep track of
previous events, extensive number of emotions to be represented) and the lack of context
handling of the OCC model when creating believable characters. Nonetheless, Ortony
14
2.1. PSYCHOLOGICAL THEORIES OF EMOTION
and colleagues presented their awareness about these issues, and they stated that the OCC
model is a basis model to define human emotions with a cognitive and individual approach.
Therefore, the problems that Bartneck explains are details that need to be handled in a
separate way or as a component of the OCC model.
Scherers Model
A model that does not deal with categories but with processes is the Component Process
Model (CPM), proposed by Klaus Scherer [140]. The CPM is a dimensional dynamic
model that defines emotions as adaptive reactions to events driven by processes of the
organism, which consists of five components corresponding to five distinct functions: (1)
Cognitive: evaluation of objects and events; (2) Peripheral efference: system regulation, (3)
15
CHAPTER 2. PSYCHOLOGICAL THEORIES
Frijdas Theory
Nico Frijda proposed an appraisal theory of emotions based on the term concern. According
to Frijda, a concern is what gives a particular event its emotional meaning. Emotions arise
in response to events that are important to the individuals goals, motives, or concerns [60].
For him an emotion is defined by six characteristics that describe its function: (1)
Concern relevance detection, (2) Appraisal, (3) Control precedence, (4) Action readiness
changes, (5) Regulation, and (6) Social nature of the environment. On the other hand, the
emotion process can be described with three lines: the core process (leads from stimulus
event to response), the regulation line (processes that intervene in the core process) and
the line of inputs other than the stimulus event. The outputs are: the overt response and
physiological changes [59]. It is depicted in Figure 2.4.
The advantage of Frijdas model is that it can be formalized in such a way that form
the basis of a computational model, as used in the architecture of a computer agent [109].
16
2.2. PSYCHOLOGICAL THEORIES OF PERSONALITY
Figure 2.4: Frijdas emotion process. Adapted from Fig. 9.1. in [59].
Concerning personality, theories happen to be very different among them. The state of the
art theory is the Five Factor Model, or Big Five, which proposes five almost independent
dimensions providing a very clear definition of personality. Nevertheless, it is not clear
(psychologist are still doing studies) how these dimensions can be related between each
other. Another theory based on FFM factors is the AB5C model, proposed by De Raad.
The advantage of this theory is that it allows to combine two factors, obtaining all the
corresponding adjectives needed to define a characters mixed personality.
17
CHAPTER 2. PSYCHOLOGICAL THEORIES
The traits were found in self-reports and ratings, in natural languages and theoretically
18
2.2. PSYCHOLOGICAL THEORIES OF PERSONALITY
based questionnaires, in children, college students, and older adults, in men and women,
and in English, Dutch, German, and Japanese samples. All five factors were shown to
have convergent and discriminant validity across instruments and observers, and to endure
across decades in adults [97]. It provides the model with two advantages: universality and
applicability.
Regarding universality, the FFM is strongly rooted in biology, and it has been found
that each of the five factors is heritable [98]. Regarding applicability, the FFM can be
used in different branches of psychology: industrial, organizational, clinical, educational,
forensic, and health psychology. Another advantage of the model is that any personality
type can be represented through the combination of the five traits, because they are found
to be independent from each other.
Although the FFM is the most used personality model to date, some psychologists
criticize the methodology, and the number of traits. Some say that five factors are too
many factors, but studies demonstrate that five factors are just right. Others say that
five factors are insufficient to summarize all that we know about individual differences
in personality. To this respect, the authors reply that they merely represent the highest
hierarchical level of trait description.
The motivation for circumplex models is that they provide much more opportunity for
identifying clusters of traits that are semantically cohesive.
Wiggins Model
Wiggins et al. [155] reoriented the Interpersonal Circumplex, or IPC, which defines a broad
set of interpersonal traits that are directly related with affective and cognitive behavior.
The IPC has sixteen dimensions that were reduced to eight, where each octant is a combi-
nation of the dimensions: Dominance/Passiveness and Affect/Hostility. The octants that
are adjacent to each other share attributes and the ones that are opposed are inversely
related. Figure 2.5 shows the circumplex.
Traditionally this circular structure has been used to define interpersonal rela-
tions and to explain users trustworthiness in collaborative virtual environments and
telemedicine [23]. The advantage of this model is that allows a fine-grained definition
of personality traits.
19
CHAPTER 2. PSYCHOLOGICAL THEORIES
AB5C Model
The Abridged Big Five Dimensional Circumplex (AB5C) [74] taxonomy of personality
traits consists of 10 circumplexes that were obtained by the pair-combination of FFM
traits. Hofstee et al. found that by blending the FFM traits by pairs, a much tighter
conceptual structure that seems to work in practice was achieved. On the other hand,
the model is less restrictive than simple-structure models and two-dimensional circumplex
models, like the Wiggins model [155]. Figure 2.6 shows one of the ten circumplexes that
combines Extraversion or Surgency (Factor I) with Emotional Stability (Factor IV).
De Raad and Barelds [128] used two factors, Virtue and Dynamism, to organize the Big
Five variables in a circumplex model. The advantage of using this model is that the
positions of the trait-variables relative to each other become clear. This organization is
shown in Figure 2.7, where variables can be read focusing on two sets of opposite clusters.
20
2.2. PSYCHOLOGICAL THEORIES OF PERSONALITY
Figure 2.6: AB5C - Extraversion (I) and Emotional Stability (IV) (Fig. 1. from [74])
21
Figure 2.7: Circumplex - Two-factors: Virtue and Dynamism (Fig. 8.4. from [127])
CHAPTER 2. PSYCHOLOGICAL THEORIES
2.3.1 Ekman
Ekman [43] distinguished mood from emotions in terms of their time course (moods last
for hours or days) and of what should be found in the neural circuitry that directs and
maintains each of these affective states.
Most of the times, laypeople use the same word to refer to a mood or to an emotion.
For instance, the word irritable would mean low-intensity anger or a long-lasting state.
Another criterion is that moods seem to lower the threshold for arousing emotions; as if
the person is seeking an opportunity to indulge the emotion relevant to the mood.
Ekman also stated that moods do not have their own unique facial expressions, while
emotions do. Another characteristic of mood is that usually people cannot recall what
situation brought them to a certain mood, while they can do that with emotions. Internal
chemical changes can also change the mood, for instance, lack of sleep or food.
Albert Mehrabian proposed a framework for definition and measurement of different emo-
tional states, emotional traits, and personality traits in terms of three nearly orthogonal
dimensions: Pleasure, Arousal, and Dominance; which define the PAD Space.
There are two PAD Scales. One for definition of emotional states, or emotions (PAD
Emotion Model); and the other for definition of temperament (PAD Temperament Model).
Both the PAD Temperament Model and the PAD Emotion Model allow us to predict the
correlation between any two traits (temperament) or states (emotions) for which PAD
components have been experimentally identified. In this way, an agent would be infused
with personality characteristics, or emotions, that appear to have life-like quality. For
example, an agent that is configured to be neurotic would thus manifest related character-
istics (e.g. anxiety, proneness to binge eating, depression, or even panic disorder). On the
22
2.3. PSYCHOLOGICAL THEORIES OF MOOD
other hand, based on the correlation among traits, this neurotic agent would not be likely
to exhibit extroverted or nurturing traits [103].
During the past few years this model has been used in computational model for repre-
sentation of mood in virtual characters [4], [63], [80]. The reason for using a temperament
model as a mood model is given by the fact that in the PAD space a set of different affective
values are produced depending on the values of pleasure, arousal and dominance, which
in turn change over time. As the combination of these three dimensions produce eight
different octants, then we can assume that these octants are moods.
In [105], Emotional States refer to transitory conditions of the organism (e.g. feeling
alert vs. tired, happy vs. unhappy), which can be seen as emotions and/or moods. Tem-
perament refers to an individuals stable emotional characteristics (i.e. emotional traits
or generalized emotional predispositions). More precisely, temperament is an average of
the states of pleasure, arousal, and dominance across representative life situations.
The UWIST Mood Adjective Checklist (UMACL) is a tool for measuring mood. Matthews
et al. [96] defined mood as an emotion-like experience lasting for at least several minutes.
Some of the previous mood models they studied to obtain their final scale were the
one proposed by Mehrabian and Russell [106], which used three bipolar factors: pleasure-
displeasure, arousal, and dominance-submissiveness; Zevor and Tellegens [157] two factors-
model: positive affect and negative affect; Thayers [145] that also obtained these two fac-
tors, but he labeled them energetic arousal and tense arousal ; and Mackay et al.s [90]
who identified bipolar dimensions related to hedonic tone or feeling of pleasantness-
unpleasantness, and arousal.
In the end, Matthews et al. proposed a three-dimensional model of mood: energetic
arousal, hedonic tone, and general arousal. It is of great importance in clinical research,
because of its apparent ability to discriminate between depressed (low hedonic tone) and
anxious (high tense arousal) mood states.
23
CHAPTER 2. PSYCHOLOGICAL THEORIES
as feelings of nervousness, sadness, irritation, and guilt. The Positive Affect reflects the
experiencing of some type of positive mood as feelings of joy, energy, enthusiasm, and
alertness.
The model classifies positive and negative moods in four basic types: high positive/low
negative (e.g. feeling happy), high positive/high negative (e.g. mixture of fear and ex-
citement in a roller coaster), low positive/high negative (e.g. feeling depressed), and low
positive/low negative (e.g. disengaged state while watching television).
Watson and Clark [152] developed their own mood inventory named PANAS-X, which
is an extension of the original PANAS (Positive and Negative Affect Schedule). It consists
of 11 scales that assess specific types of affect: 4 basic negative affects (fear, sadness, guilt
and hostility), 3 basic positive affects (joviality, self-assurance and attentiveness), and 4
other affective states (shyness, fatigue, serenity and surprise).
2.4 Summary
In this chapter we have reviewed some psychological theories of emotions, mood and per-
sonality that are relevant in the field of Affective Computing. From Darwin, who studied
the universality of facial expressions, to Thomas Cochrane, who proposed a novel theory for
the implementation of appraisal, several psychologists have come up with different ways to
study emotions. Categorical, dimensional, and appraisal models of emotions are the three
types we have overviewed, being the OCC model the most used in computational models
to date. Regarding personality, the Five Factor Model, or Big Five is still the state-of-art
personality model. One reason is its replicability along different studies. Another reason is
that its five dimensions allow the description of any type of personality. Finally, the study
of mood is becoming more and more important in the Affective Computing field. There-
fore, efforts are being directed to its representation in computational models, going from
bi-dimensional representations (good mood and bad mood) to the 8-moods PAD Space.
24
Chapter 3
Dr. Walter Gibbs: Ha, ha. Youve got to expect some static. After all, computers are just
machines; they cant think.
Alan Bradley: Some programs will be thinking soon.
Dr. Walter Gibbs: Wont that be grand? Computers and the programs will start thinking
and the people will stop.
TRON (1982).
Thanks to the efforts in the fields of Affective Computing, Artificial Intelligence, Computer
Graphics, and Cognitive Sciences, the creation of virtual characters has been improved and
enriched through the years.
Some researchers have proposed computational models based on psychological theories
to elicit different affective traits and behaviors in the characters. Others have focused on
studying which behaviors are perceived as manifestation of different affective traits. While
the first ones aimed for a character that feels and react accordingly to those feelings; the
last ones aimed for cues that make a character looks like feeling.
In the following we will review previous works grouped by those that propose com-
putational models of affect for Embodied Conversational Agents (ECAs), and those that
studied the perception of affect in the face and head. Finally, a summary of the section is
provided.
25
CHAPTER 3. RELATED WORK IN AFFECTIVE COMPUTING
3.1.1 Cathexis
Juan Velasquez [147] presented Cathexis, one of the first distributed, computational models
that represented the dynamic nature of emotions, moods and temperaments, as well as their
influence on the behavior of synthetic autonomous agent. The architecture of the model
presented two components: the Emotion Generation System and the Behavior System.
The emotion generation system used appraisal theories with other emotional theories
based on physical reactions. The implementation was based on proto-specialists agents pro-
vided with sensors to monitor internal and external stimuli, allowing the elicitation of family
of emotions (e.g. Fear, Fright, Terror, etc.). Emotions could be basic or blended/mixed
(e.g. Grief, a combination of sadness and anger).
Moods were defined from a psychobiological perspective as levels of arousal that in-
fluence the activation of emotions. Temperaments were different values of thresholds that
controlled the intensity and arousal of emotions. To compute the intensity of emotions,
Velasquez took into account its previous level of arousal, the contributions of each emotion
elicitor for that emotion, and the interaction with other emotions.
The behavior system decided which behavior to display given the agents emotional
state. Each behavior contained two major components: one for generation of prototypical
facial expressions, body postures and vocal expressions; and other for identification of
motivations for behaviors and action tendencies (e.g. fighting, insulting, biting, etc.).
The system was implemented in an object-oriented framework. The ECA was Simon, a
synthetic agent representing a baby. The users interacted with Simon through an interface,
providing external stimuli that caused him to react emotionally. Our model is similar to
Cathexis in the sense that we also took into consideration internal and external stimuli.
The difference is that our external stimuli is provided by events happening inside the
26
3.1. COMPUTATIONAL MODELS OF AFFECT
virtual world, and we do not consider physiological elements to elicit or manifest affective
phenomena. Figure 3.1 shows the facial expressions of Simon.
27
CHAPTER 3. RELATED WORK IN AFFECTIVE COMPUTING
expressing facial emotion content with speech, and in some cases, music [49]. The main
difference of our model is that we do not represent physiological responses. Elliott modeled
mood as the factor that changed thresholds for emotion activation, while we used the PAD
model [105] to represent it. Finally, memory was a factor that Elliott took into consider-
ation using databases, while we relied on ontologies to reuse already existent knowledge,
but so far we do not deal with memory.
Andre et al. [3] presented one model that integrated personality and emotions to create
interactive virtual characters. In the Puppet project, children were intended to gain a basic
understanding on how emotional states change, as well as to comprehend how physical and
verbal actions in social interaction can induce emotions in others.
The architecture considered deliberative planning (goals) and reactive plans (inten-
tions), built on a BDI framework. They considered a knowledge base (database that
contains the world model), a plan library (collection of plans to be used by the agent to
achieve its goals), an intention structure (internal model of the current goals or desires,
and instantiated plans or intentions), and an interpreter (resolves conflicts, select a plan
and execute it). Events might be elicited from the virtual environment, or from the user
input. They also introduced body states (hunger, fatigue, boredom).
The modeled emotions were: anger, fear, happiness and sadness. These could be elicited
through OCC rules, or by the child interacting with the system. Regarding personality,
they considered two traits from the FFM, extraversion and agreeableness. Interaction could
be performed in three ways: the child controlled one avatar and interacted with others,
the child observed the interaction of the avatars, and the child was like the director of the
theater controlling the behavior of all characters.
Visualization was done through 2D cartoon-like characters that form part of a farm: a
farmer and a animals. We chose to explain the Puppet project because as it did, we are
considering a model of the world, also named knowledge base, which is modeled through
ontologies. The main difference is the planning behavior, which we do not consider, because
we are interested in visualizing affect through facial expressions.
28
3.1. COMPUTATIONAL MODELS OF AFFECT
29
CHAPTER 3. RELATED WORK IN AFFECTIVE COMPUTING
3.1.5 Greta
Rosis et al. [35] designed and prototyped Greta, a 3D Embodied Conversational Agent
(ECA) provided with Mind and Body to enhance in the user the impression of communi-
cating with a specific person.
The mind of Greta was designed based on a BDI (Belief-Desire-Intention) model [130]
considering: temperament and personality, social context, dynamics of the agents state,
response decay, and multiple emotions. It means that her mind has a representation of
the beliefs and goals that elicit emotions and the decision to display them or not. They
also combined emotions and considered intensity changes with time, and how each of them
prevails according to the agents personality and social context of the conversation.
The body of Greta used a repertoire of signals to be employed during communication
like facial expressions, head movements or gaze direction. In recent versions of Greta [93],
the agent also produced gestures (arms and hands movements) and upper body movements.
To implement Gretas mind, they used Dynamic Belief Network (DBN). To keep Gretas
mind independent of her body, they defined a mark-up language (Affective Presentation
Markup Language - APML [34]) to associate semantics to the natural language utterances.
One of the advantages of this system was its multimodality and domain-independence.
By not using emotional and personality models they built a fine-grained cognitive structure,
in which the appraisal of events was represented in terms of the agents system of beliefs
and goals. The problem arose with the use of DBNs, because the number of nodes increases
considerably with the number of modeled emotions. The difference with our model is the
implementation of the mind of our system, which was done through ontologies, and the
interaction between affective phenomena that, in our case, was done using the PAD Space.
Figure 3.3 shows the first attempts on creation of facial expressions using Greta.
Figure 3.3: (A) Gretas Fear and (B) Gretas Joy (Fig. 36. and 37. from [122])
30
3.1. COMPUTATIONAL MODELS OF AFFECT
Figure 3.4: Facial Expressions for Anger, Surprise and Sadness (Fig. 7. from [41])
31
CHAPTER 3. RELATED WORK IN AFFECTIVE COMPUTING
3.1.7 ALMA
Gebhard [63] proposed ALMA (A Layered Model of Affect) to simulate the interaction of
emotions, mood, and personality. The model used OCC emotions and the FFM personality
model. These elements were mapped into the PAD Space [104] to represent moods.
Elicited emotions were represented in the PAD space as vectors and were condensed in
a center of mass, influencing the mood and making it jump from one octant to the other
of the PAD model. Personality was defined as a default mood using a set of equations
provided by Mehrabian [105]. It influenced emotions by giving an initial offset to those
related to the activated personality traits. Finally, mood updated personality, using the
inverse of the equations employed to compute the default mood.
ALMA was implemented using AffectML, an XML-based affect modeling language. A
characters personality profile consisted of the personality definition and subjective ap-
praisal rules. At runtime, the affect computation periodically updated the affective profile
of all characters, appraised relevant input for all characters, and gave as output a set of
emotion eliciting conditions. These were used to update a characters emotions and mood.
Regarding visualization, facial expressions, or visual cues as blush of shame showed
the experienced emotion. Idle behaviors represented the mood. Nevertheless, ALMA has
been mainly used for dialog-generation in virtual characters [64]. Our work is based on this
model because it provided the framework to interrelate personality, mood and emotions
in a novel and practical way. Figure 3.5 shows the PAD space where affective traits are
mapped and the resultant behavior in a character.
3.1.8 FATIMA
Inspired by the work of traditional character animators, Dias and Paiva [38] proposed an
architectural model to build autonomous characters whose behavior was influenced by their
emotional state and personality.
The architecture presented two layers: the reactive layer (hardwire reactions to emo-
tions and events that must be rapidly triggered and performed after the appraisal process),
and the deliberative layer for the agents planful behavior. FATIMA generated characters
behavior based on appraisal and coping processes. Appraisal focuses on the goals of the
character, triggering the emotions to take into account to prepare a plan. Coping de-
pended on the emotional state and personality of the character. They considered two
types of coping: problem focused (set of actions to achieve and execute a result) and emo-
32
3.1. COMPUTATIONAL MODELS OF AFFECT
Figure 3.5: (A) Character Behavior ([63]). (B) Shame representation over time ([64])
They used the OCC model for emotions, and also represented arousal (degree of ex-
citement of the character) and mood. Mood, represented as good or bad, was an overall
valence of the characters emotional state and influences the intensity of emotions. Person-
ality was defined by: a set of goals, a set of emotional reaction rules, the characters action
tendencies, emotional thresholds, and decay rates for each emotion type.
FATIMA provided synthetic characters that were believable and empathic. They had
cognitive capabilities, interacted with external users and were domain independent. This
model was mainly goal-based, letting motivations and standards to define the characters
personality. Our model is a simplified version of a goal-based model, which takes it into
account to generate emotions, but also considers preferences and admiration for other
agents to elicit other OCC emotions.
FATIMA was implemented in the computer application FearNot! [37], which was de-
veloped to tackle and eventually help to reduce bullying problems in schools. Figure 3.6
shows a screenshot of FearNot!.
33
CHAPTER 3. RELATED WORK IN AFFECTIVE COMPUTING
Figure 3.6: (A) Characters in FearNot! and (B) Facial Expressions (Fig. 3 and 4 [121])
3.1.9 WASABI
Becker-Asano [12] presented WASABI Affect Simulation for Agents with Believable
Interactivity, a computational simulation of affect for embodied agents. The architec-
ture of WASABI distinguished two layers: a physis layer and a cognition layer; and three
affective states: mood, primary emotions, and secondary emotions.
In the physis layer, primary emotions were produced from the non-conscious appraisal
of the stimuli. The set of primary emotions were anger, annoyance, boredom, concentration,
depression, fear, happiness, sadness, and surprise.
In the cognition layer the agent used its BDI-based cognitive reasoning abilities to
update his memory and generate expectations. From the conscious appraisal, secondary
emotions were produced. They were first filtered in the PAD space influencing the embodi-
ment of the agent and the expression of these emotions. Secondary emotions corresponded
to the prospect-based emotions OCC-cluster: hope, fears-confirmed, and relief.
Mood was the background state that influenced the elicitation of emotions. In contrast
to ALMA (Section 3.1.7), mood was not derived from PAD space, but modeled as an
agents overall feeling of well-being on a bipolar scale of positive vs. negative valence.
The ECA where WASABI was tested was Max. He was employed in a museum applica-
tion, where he conducted multimodal smalltalk conversations with visitors. Furthermore,
WASABI has also been applied to a gaming scenario, in which secondary emotions were
simulated in addition to primary ones [13]. The main difference of WASABI with our model
is the differentiation between primary and secondary emotions at the appraisal level. We
34
3.1. COMPUTATIONAL MODELS OF AFFECT
make this distinction when visualizing facial expressions of our characters, and moreover,
our secondary emotions are mixtures of primary emotions. Although we also use the PAD
space, we consider mood traits as the octants of the space instead of limiting them to
positive or negative. Figure 3.7 shows the virtual agent Max.
3.1.10 EMA
Marsella and Gratch [94] [95] proposed EMA (EMotion and Adaptation), a computational
framework that represented the dynamics in appraisal, which make the elicited situations
change based on inferences and previous knowledge to cope with that situation.
The agents interpretation of its agent-environment relationship was called causal in-
terpretation. It provided a explicit representation of the agents beliefs, desires, intentions,
plans and probabilities used for the appraisal processes. This causal interpretation changed
in time depending on the agents future observations or inferences. Regarding events, they
were defined in terms of appraisal variables which were: Perspective, Desirability, Likeli-
hood, Causal Attribution, Temporal Status, Controllability and Changeability.
Coping refers to how the agent reacts to the appraised events. Coping strategies in
EMA used the cognitive operators of the appraisal process, and decided for the most
suitable actions to be performed. Strategies included: planning, seek instrumental support,
procrastination, positive reinterpretation, acceptance, denial, mental disengagement, shift
blame, seek/suppress information.
As AR (Section 3.1.2) and FATIMA (Section 3.1.8), EMA followed Frijdas theory, so
then appraisal and mood elicited emotional states. Finally, coping response was biased by
this overall mood state. Regarding knowledge, it was represented through propositions,
35
CHAPTER 3. RELATED WORK IN AFFECTIVE COMPUTING
and SOAR [114] was used to model cognitive operators (e.g. update belief, output speech,
listen to speaker, initiate action, and so on) [95].
The relevance of this work is the single layered appraisal model, which resulted simply
enough to be implemented in virtual characters, so they can make congruent inferences
about their world and cope accordingly to them. We also followed Frijdas background,
because through our appraisal mechanism (ontologies) we elicit emotional states. We also
used variables as desirability, likelihood and temporal status as the variables to appraise
the different situations.
Kasap et al. [80], proposed a memory-based emotion model that intended to achieve a more
natural interaction between the user and a virtual character. Through face recognition
techniques, the character could remember a users face and automatically adjust the
current interaction on the basis of its existing relationship with the user.
To model emotions Kasap et al. used the reduced OCC model, which has 12 emotions
(six positive: joy, hope, relief, pride, gratitude, and love; and six negative: distress, fear,
disappointment, remorse, anger, and hate). In addition, they used another 4 user-related
emotions: happy-for, gloating, sorry-for, and resentment. Personality was modeled using
FFM and moods were represented in the PAD space, defined by equations proposed by
Mehrabian [105].
This architecture was similar to ALMA (Section 3.1.7), but Kasap et al. integrated
the interpersonal-relationship concept in which emotion, mood, personality, and social
relationships affect to each other. Long-term memory allowed the virtual character to
store specific interaction sessions, and then retrieving this information as needed.
At the beginning, during, and at the end of each interaction, emotions and moods were
updated using the values provided in a matrix that related emotions (rows) with moods
(columns). Relationships were framed on the dimensions of Dominance and Friendliness,
following Argyles model [5]. Long-term memory was represented through structures named
episodic memory, which kept track of people that had interacted and calculated relationship
levels with them [91].
Their example application was Eva, a geography teacher who had a good and a bad
interaction with two different students, showing the differences in the interactions outputs.
Figure 3.8 shows the facial expression of Eva during interaction.
36
3.1. COMPUTATIONAL MODELS OF AFFECT
This model is very similar to our model, regarding mapping of affective traits in the
PAD space. While they consider relationship during the calculation of the affective state,
we take it into consideration while appraisal of the event (in the ontology). Also, Kasap
et al. only modeled facial expressions of emotions, while we map moods pleasure, arousal
and dominance values in the face.
3.1.12 OSSE
Ochs et al. [118] proposed a model of the dynamics of social relations, based on emotions,
for Non-Player Characters (NPC). NPCs are defined according to their personality traits
and social roles (relation with other NPCs).
To represent the world surrounding the NPCs and its events, they used a formal rep-
resentation that included: a vocabulary ( entities with no reasoning capacity and actions),
NPCs attitudes functions (towards objects and other characters), praise functions (to-
wards actions), events as the 4-uplet <agent, action, patient, degree certainty>, emotion
representation, emotions intensity (degree of desirability, of praiseworthy/blameworthy),
personality traits, dynamics of social relations as the 4-uplet <liking, dominance, familiar-
ity, solidarity >, and social roles (e.g. employee/manager, child/father).
In the architecture, events were triggered by the scenario and appraised using OCC-
based algorithms. It elicited emotions that were influenced by personality. In turn, emo-
37
CHAPTER 3. RELATED WORK IN AFFECTIVE COMPUTING
tions affected the NPCs emotional state and altered the social relations with other NPCs.
In the implementation of the architecture, emotion representation was done through
the 10-emotions reduced OCC model [119], and they just considered the personality traits
Neuroticism and Extraversion. For instance, a more extroverted NPC experiences higher
joy, hope, pride and relief emotion values. Emotional decay functions and emotion thresh-
olds depended on the personality. To test the model, they used interactive dialogs that
simulated scenarios with different contexts (e.g. police interrogation and job interview).
This model is important because it focused on the dynamics of emotions, and not
just in their representation. They considered most of the aspects that participate in the
emotion elicitation from a cognitive perspective, achieving a simple representation. Ochs et
al.s model conceptually resembles our model because actions, objects, agent, preferences,
and events are related in a similar fashion. Nevertheless, we make use of ontologies to
represent the knowledge that they implemented through functions. In this way, we reuse
existent knowledge and perform inferences about the context defined. Moreover, they did
not consider emotions as happy-for (sorry-for ), which occurs when someone we appreciate
undergoes desirable (undesirable) events; nor gloating (resentment). In our model we can
elicit these emotions because we have a liking level for another agent which is represented
in the ontologies.
Courgeon et al. [31] proposed a model based on Scherers Componential Process Model
(CPM) (Section 2.1.3). Their motivation was the appraisal of affective events in real-time.
For this reason, they used Scherers model, which defines appraisal based on the evaluations
an organism make in order to survive.
MARCs architecture had an appraisal module that evaluated the elicited events and
generated the parameters for facial animation. They just used 7 out of the 10 appraisal
sub-checks of the CPM: expectedness, unpleasantness, goal hindrance, external causa-
tion, copying potential, immorality, and self consistency. For each sub-check they defined
Gaussian curves, based on the proposed discrete prediction values (open, very low, low,
medium, high, and very high), because they provided a linear quantification of discrete
values. Therefore, they could evaluate the sub-checks in a continuous scale.
The process can be summarized as, when an event occurs the relevant emotions for that
event are evaluated. They used a set of four emotions: joy, sadness, anger, and guilt. For
38
3.1. COMPUTATIONAL MODELS OF AFFECT
each emotion, they multiply its relevance by the emotions Gaussian value of all its sub-
checks. Then, the final emotion is set with a value and animation parameters are generated.
Visualization was done using MARC, a virtual character that displayed sequential facial
expressions. Figure 3.9 shows the systems architecture.
We decided to review this work because it is a novel appraisal model that uses the
CPM in a simple manner. In essence, Courgeon et al.s model follows our own framework
organization: we have an event elicitor module that performs appraisal, then a module
where the affective parameters are fine-tuned and values for facial animation are generated,
and finally a facial animation module.
39
40
Model Affective Elements Psychological Implementation Visualization ECA
Background
Cathexis - 6 basic emotions (Ekman) Appraisal + Emotions: Sensor agents - Basic facial Simon
(1997) - Mixed Emotions Physical motivations (neural, sensorimotor, expressions
- Mood (Roseman model) motivational, cognitive). - Body postures
- Temperament (emotions Behavior: motivations, - Vocal expressions
thresholds) external stimuli.
Affective - 24 emotions (OCC model) BDI Databases for expectations, - Facial expressions Talking heads
Reasoner - Mood concerns-of-other, goals, (either computer
(1998) standards and preferences or human)
(GPS). OCC rules, actions,
conflicts sets.
PUPPET - 4 Emotions (anger, fear, Deliberative and OCC rules - Facial expressions Farm
(1999) sadness, happiness) reactive planning + - Sounds (e.g. cat characters
- Personality (extraverted BDI purrs and hisses) (farmer,
and agreeableness) pigs, etc.)
Multilayer - 24 emotions (OCC model Emotional reactive Bayesian Belief Networks - Basic facial Virtual character
Personality + Surprise + Disgust) (BBF) for personality. expressions
Model - Personality (FFM) Probability transitions - Vocal expressions
(2002) - Mood (good/bad/neutral) matrices for mood.
Greta - 24 emotions (OCC model) BDI Dynamic Belief Networks for - Facial expressions Greta
(2003) emotion representation. - Body postures
- Head movements
and gaze
- Vocal expressions
Generic - N emotions Appraisal (OCC Matrices of influence. - Basic facial not found
Personality - P Personalities model + personality) Functions to compute expressions
and Emotion - M Moods affective traits changes. - Vocal expressions
Model (2004)
FATIMA - 24 emotions (OCC model) Appraisal and Reactive and deliberative - Facial expressions FearNot!
(2005) - Personality (goals, Coping appraisal. - Body postures (John the victim,
emotional reaction rules, Problem focused and Martinha the
action tendencies, emotional emotional focused coping. neutral and Luke
thresholds and decay rates) the bully)
- Mood (good/bad)
CHAPTER 3. RELATED WORK IN AFFECTIVE COMPUTING
ALMA - 24 emotions (OCC model) Cognitive Appraisal Appraisal rules: basic, act, - Facial expressions Cyberella,
(2005) - Personality (FFM) emotion and mood display - Body postures Cross Talk,
- Mood (PAD space) (AffectML) - Written dialog VirtualHuman,
Eric, Virtual
Poker-Character
WASABI - Primary emotions BDI + bodily Physis and cognitive layers - Facial expressions Max
(2008) (prototypical emotions) dynamics in PAD Space. - Vocal expressions
- Secondary emotions
(prospect-based emotions
in OCC-cluster)
- Mood (positive/negative)
EMA - Emotions (surprise, hope Dynamic Appraisal Cognitive architecture not found not found
(2009) joy, fear, sadness, anger and coping (SOAR)
and guilt)
- Mood (aggregated
emotional states)
Memory-based - 12 Emotions Cognitive appraisal Matrices relating - Facial expressions Eva
Emotion (reduced OCC) + Interpersonal emotions and moods. - Body postures
Model (2009) - Personality (FFM) relationship Probabilities of - Written dialog
- Mood (PAD space) (dominance- affective state change.
3.1. COMPUTATIONAL MODELS OF AFFECT
41
CHAPTER 3. RELATED WORK IN AFFECTIVE COMPUTING
42
3.2. VISUAL PERCEPTION OF AFFECTIVE PHENOMENA
Krahmer et al.[83] investigated three potential personality cues: gaze, speech and eye-
brow movements, which were assigned variants of Extraversion and Introversion. The
novelty was the study of the perception of these cues combined in different personality
profiles (e.g. introverted gaze, introverted speech and extraverted brow). They used a
virtual talking-head with the three variants of personality cues, which had to be rated by
subjects while reading a poem in Dutch. The results showed that extravert brows were
perceived more extraverted when combined with extraverted gaze, and introvert brows
were perceived as more introvert with introvert gaze. Concerning combinations of cues
they found that including an extravert feature in an introvert agent does not imply that
subjects perceive that agent as more extraverted. They also observed that the personality
profile of the character had no influence in the quality assessment. They also confirmed
that speech influences perception of personality.
Arya et al. [7] experimentally associated facial actions (head tilting, turning, and nod-
ding, eyebrow raising, blinking, and expression of emotions) and their frequency and du-
ration to the dimensions of Dominance and Affiliation, proposed by Wiggins [155]. Their
objective was to have an agent with certain facial actions that can be perceived by the
user as certain personality types, instead of establishing the relation between personality
types and facial actions. The results showed that joy was highly related to the Affiliation
dimension and was seen more dominant. Contempt was also seen as very dominant but
correlated negatively with Affiliation. Slowly and fast raising of both brows was seen as an
Affiliation signal. Slowly turning the head was seen as very dominant, while slowly moving
the head to the side was seen as very submissive. On the other hand, raising fast one brow
was seen as very dominant, while a fast avert gaze was seen as more submissive.
Zammitto et al. [156] tried to expand the personality model proposed in Arya et al. [7]
by using the AB5C model (Section 2.2.3), which provided the framework for going from
mapped circumplexes to meticulous use of labels. As a result they obtained adjectives for
identifying the Big Five personality factors, plus specific visual cues for two of them. Their
idea was to incorporate visual cues for the rest of the Big Five traits.
Bee et al. [15] analized how facial displays, eye gaze, and head tilts express social
dominance. Using one-way ANOVAs and two-tailed t-test, depending on what they wanted
to obtain, they obtained results as the emotional displays of joy, anger and disgust were
perceived as high dominant, while fear was less dominant. Regarding gaze, joy, disgust and
the neutral display with direct gaze (looking straight to the user) were the most dominant,
while fear was again the less dominant. Nevertheless, anger with averted gaze was perceived
43
CHAPTER 3. RELATED WORK IN AFFECTIVE COMPUTING
as high dominant. Concerning influence of head orientation, significant effects were only
obtained for anger, sadness and the neutral display.
McRorie et al. [101] proposed a model to construct Sensitive Artificial Listeners (SAL)
using Eysencks model of personality to generate behavioral characteristics (e.g. visual
cues) from it. The idea was to link verbal/non-verbal behavior with impressions of per-
sonality. In order to do that, first they defined the agents specifying its characteristics and
mental state: agree, disagree, accept, refuse, believe, disbelieve, interest, not interest, like,
dislike, understand, and not understand. Then, they computed the agents behavior when
it was listening to the user, selecting the most appropriate signal. Finally, they determined
the corresponding non verbal behaviors taking into account the characters modality pref-
erence and expressiveness parameters. For example, extraverts tend to demonstrate more
body movements, and display greater levels of facial activity and gesturing, more frequent
head nods, general speed of movement, and direct facial posture and eye contact is more
likely to be maintained. For instance, their character Spike, who is angry and argumen-
tative, conveys negative communicative functions, in particular dislike, disagreement and
lack of interest.
Cloud-Buckner et al. [28] conducted an experiment to investigate how people perceived
personality on avatars based on actions, language, and behavior; and if race and gender of
the avatar influenced this perception. The independent variables taken into consideration
were race (dark or white), gender (male or female), and personality (P1: friendly, outgoing,
self-sufficient, with high-activity level and some anger; P2: introverted, self-conscious,
cooperative, orderly, modest, disciplined and sympathetic to others). Results showed that
race and gender of the avatar did play a role in the perception of personality. For example,
for P1 and P2, when there was a difference on Race, the dark skin had a higher rating for
every item except anger. When there was a difference on Gender, in P1, the male had a
higher score, but in P2, the women had a higher score.
Neff et al. [113] studied the impact of verbal and non-verbal factors in the perception of
the Big Five trait of Extraversion. The non-verbal expression of Extraversion was given by
the head and body attitude: gesture amplitude, direction, speed, and movement of body
parts. Neff et al. performed evaluation on videos of a female avatar making a restaurant rec-
ommendation and asking subjects to judge her extraversion level and naturalness. Results
showed that linguistic extraversion has more influence on the perception of extraversion
than gesture rate or gesture performance; that indeed higher gesture rates are perceived as
more extraverted; changes in gesture performance correlate with changes in extraversion
44
3.3. SUMMARY
perception; and the combination of higher gesture rates with a more extraverted gesture
performance increase the perception of extraversion.
Bee et al. [16] examined the interaction of verbal and non-verbal behavior to create
the impression of dominance in intelligent virtual agents. Non-verbal communication is
achieved by gaze and head movement, which is associated with social dominance; and
gestures that acquaint friendliness and enthusiasm. Verbal communication is achieved
through linguistic behavior, which is associated with the personality traits of extraver-
sion and agreeableness. Their novel result demonstrated that linguistic personality traits
(extraversion and agreeableness) influence the perception of dominance. Also, gaze-based
dominance influences the perception of personality traits.
3.3 Summary
This chapter can be summarized as the overview of models to make character feel and
to make character look like feeling. In the first group we found the models of Velasquez,
Andre et al., Kshirsagar et al., Egges et al., or Gebhard et al., which have become State
of Art when studying computational models of affect. Other works as Kasap et al., Ochs
et al., and Courgeon et al., propose new models that intend to improve the creation of
affective characters. Table 3.1 shows the highlights of the reviewed models as well as of
our model.
In the second group we have briefly described studies related with the perception of
personality and their relation with different visual cues. This studies have been mostly
performed by psychologists, but recently researchers in Computer Science like Arya and
colleagues, Zammitto and colleages, or Bee and colleagues are experimenting with the
implementation of these cues in virtual characters and explore how they enhance the per-
ception of emotions, mood and personality.
45
CHAPTER 3. RELATED WORK IN AFFECTIVE COMPUTING
46
Chapter 4
Framework Overview
Da Vinci combined art and science and aesthetics and engineering, that kind of unity is
needed once again.
Ben Shneiderman
The combination of different fields and theories to achieve believable virtual characters is
not an easy task. It could be seen in the previous chapters the amount of previous work
to give characters affective traits. It also could be seen that context is a key element in
the generation and recognition of emotions. On the other hand, little has been seen about
expression of mood or personality that could enhance the believability in the characters.
Therefore, we propose a framework that deals with the representation of context and
the visualization of emotions, mood and personality, combining ontologies, psychological
theories and previous computational models. Nevertheless, due to the novelty of this
work, every piece of the framework needs to be evaluated to prove our research hypothesis
established in the introductory chapter. Therefore, the main purpose of this framework is
to establish the basis for a model that could help in the development of applications with
storytelling, interaction with virtual characters or fast visualization of affect.
This chapter gives an overview of how we implement the necessary elements to create
believable affective virtual characters immersed in contextualized virtual worlds.
47
CHAPTER 4. FRAMEWORK OVERVIEW
As can be seen, the world provides the elements to represent the context, which are
appraised and elicits emotions that are interpreted by the affective model. Then, the
affective model combines these emotional responses with the mood and personality of the
character, to produce a final affective state, or mood. Finally, this mood is manifested by
the agent through its facial expressions. Figure 4.2 presents a more detailed schema of the
modules to be developed, and the psychological theories (or models) that will be used.
Starting from the top leftmost side of Figure 4.2, it is seen that the World is defined
in terms of the events that occur in it, and also in terms of the characters, or agents who
inhabit it. This categorization responds to idea of differentiating context regarding things
that happen outside the character (events) and inside the character (the character itself
through its psychological-affective characteristics).
An event is composed of an action that is performed, or affects a character (agent) or
an object; a time unit when it occurs; and a place where the event occurs. As for the agent,
it is defined in terms of its goals, preferences, social admiration for other agents (although
48
4.1. SYSTEM FRAMEWORK
49
CHAPTER 4. FRAMEWORK OVERVIEW
this term uses the emotional word admiration, which has a positive connotation, it can
be negative), and personality.
The World elements are inputted to the Context Representation module that is imple-
mented using two ontologies: Event Ontology, which describes all the concepts that form
part of an event; and PersonalityEmotion Ontology, which considers all the concepts that
define a character from a psychological point of view (goals, preferences, social admiration
with other agents, and personality). The output of this module is a set of emotions with
their respective intensities.
These emotions are the input of the Affective Model, which computes the interrela-
tions between emotions, moods and personality, giving more accurate affective responses.
The output of this module is a set of moods expressed in terms of pleasure, arousal and
dominance values.
The Visualization module is responsible for the generation of facial expressions, provid-
ing a visual representation for the emotions elicited in the Context Representation module
and the moods elicited by the Affective Model. Regarding personality, it does not only
influence at the computation level, but also at the visual level. That is why we explore
how some visual cues influence the perception of personality traits.
In the following sections we offer a brief description of the parameters used in each
layer of our computational affective model.
To represent the context in which the characters will be immersed we propose a Semantic
Model that will be briefly described in this section, and then explained in Chapter 5.
The internal state of a character is the result of the context appraisal, which is performed
using information as the goals, preferences, admiration for other characters, and personality
traits. Many other aspects could be taken into consideration as social role, culture, religion,
social dynamics between characters, etc. Nevertheless, we consider the former elements
the basic ones to represent the interaction of the character with his environment. In the
following subsection we will make explicit our intended meanings for them.
50
4.2. SEMANTIC LAYER: CONTEXT REPRESENTATION
Goals
In its simple sense goals define a group of affairs that an agent desires to happen. In this
dissertation, they are related to the occurrence of an event that the agent wishes to happen
and produces benefits to it. A benefit is represented as the elicitation of one, or a set of
positive emotions. Goals are considered equivalent in their structure, which means that no
difference will be made among preservation goals, partially achieved goals, and so forth.
Preferences
We have defined them as the degree of liking, or disliking, that an agent feels for sur-
rounding entities. They can be things around the character as a specific park, certain
street, flowers, chocolates, or other agents; or things that are perceived by the character
as thoughts, ideas, standards, etc. For example, the agent may be a pacifist and love the
idea of freedom, or dislike racism.
In real life people like or dislike other people to a certain degree. This feeling influences
the affective output when two characters are participating in the event, or when the event
is evaluated from the perspective of an external agent (an agent that is not participating
in the event, but has witnessed it and might feel emotions because of it).
Social role
According to Edelstein [40], roles are blend of personality traits and the work a person
does (company president, hit man, or new mother). A role emphasizes different aspects
of an individual. Although we have not set explicitly the roles as explained, we do have
established four roles which deal with the character as a protagonist or as spectator of the
event.
what: references the action that happens in the event. An action is represented with
a verb, which indicates what occurs in the event. The action has complements that
51
CHAPTER 4. FRAMEWORK OVERVIEW
complete it, giving sense to the event. Complements can be direct or indirect objects,
which are represented by Abstract Entities and Physical Entities, which were
overviewed in subsection 4.2.1.
who: indicates the role of the character with respect to the action in the event.
4.3.1 Emotions
To represent emotions in our system we chose the ones proposed in the OCC model, plus
the emotions of disgust and surprise, explained in Chapter 2. The reason to choose it,
as done previously by other researchers, is because is comprehensible and precise enough
to allow a computational implementation of emotions. These are obtained from contextual
elements for world and agents context representation, managed by the semantic layer.
4.3.2 Mood
For mood we chose the PAD Model, also explained in Chapter 2, because it links emotions,
mood and personality in the same 3D space. In PAD, each mood corresponds to an octant
of that space resulting in: Exuberant, Bored, Dependent, Disdainful, Relaxed, Anxious,
Docile and Hostile.
4.3.3 Personality
Finally, for personality, we use the traits of the Five Factor Model (FFM) [97] (also known
as Big Five) because of its proven universality after years of research using self-report ques-
tionnaires [74], and because this model can be mapped in the PAD Space with equations
provided to this end. FFM is explained in details in Chapter 2.
52
4.4. VISUALIZATION LAYER
4.5 Summary
In this chapter we provided a global view of our Computational Affective Model. It is
formed by three layers that allows the simulation of believable virtual characters.
The first layer is related with the semantics that define the context of the character
(inner context, based on a BDI theory plus personality traits; and outer context, which is
the world and its entities).
The second layer deals with the affective output produced as a result of the events
elicited by the semantic layer. In this layer emotions, mood, and personality are combined
to leave the character in an affective state defined by the values of Pleasure, Arousal and
Dominance.
The third layer is in charged of the visualization of the facial expressions that show the
affective state of the character. We have used MPEG-4 and FACS to prove that our model
is domain independent, regarding the visualization methods.
53
CHAPTER 4. FRAMEWORK OVERVIEW
54
Chapter 5
Context Representation
Appraisal can be defined as the process of understanding and evaluating what is happening
around us, allowing the elicitation of emotions. Therefore, to create emotional characters
we need to simulate a context where a variety of situations occurred and can be appraised.
This chapter is organized as follows. First, a review of previous works related with
generation of context and its representation is presented, followed by the explanation of
the implementation of our own conceptual model to represent context. Then, a practical
summary of the whole conceptual model is given, as well as a discussion about the imple-
mentation details. Afterwards, a practical example is provided to demonstrate the use of
the model. Finally, a summary concludes the chapter.
55
CHAPTER 5. CONTEXT REPRESENTATION
Hence, how a character expresses its affective state significantly depends on the context.
Without it, it might become difficult to discern between the meaning of a facial expression,
to decide how to react in certain situation, or how the course of a story/situation should
be unfolded. In this way, Kaiser and Wehrle [78] stated that facial expressions can only
be interpreted when they are inside a context (temporal or situational) that allows us to
generate them in an accurate way.
One of the problems when representing context is that aspects like culture, religion,
and other social rules and roles should also be taken into account, making the contexts
representation and evaluation an unmanageable task.
A solution for this is to represent context from a childs point of view. When we
think of children, we realize that their context has the same set of events as the one from
adults, but their appraisal lacks of social rules. Therefore, they react to those events in a
purer emotional way.
With this in mind, we decided to represent context as a child would do it, providing
a generic description of what is happening outside and inside the virtual characters, or
virtual agents - both terms will be used indistinctly. This is one of our main contributions,
because as it will be seen in the next subsection, no previous work has managed to represent
and elicit affect through ontologies.
In the field of Computer Science some areas that have work on context representation are
Affective Computing, Ubiquitous Computing, and Artificial Intelligence.
Strang et al. [141] evaluated six of the most relevant existing approaches to model
context for ubiquitous computing: key-value models, markup scheme models, graphical
models, object-oriented models, logic based models, and ontology-based models. These
approaches are also useful for any field where context representation should be taken into
account. Indeed, Strang et al. concluded that the most promising assets for context
modeling can be found in ontologies.
Ontology has been a field of philosophy since Aristotle characterized by the study of
existence, a compendium of all there is in the world. Nowadays, it has evolved in great
measure in the computer science and artificial intelligence fields [89]. An ontology is defined
as an explicit specification of an abstract, simplified view of a world to represent. It specifies
both the concepts related to this view and their interrelations [117].
56
5.1. CONTEXT - AN OVERVIEW
Krummenacher et al. [84] also stated that ontologies provide better modeling facilities
(intuitive notions of classes and properties), while being semi-structured and incorporating
a clear semantic model. They mentioned a number of upper ontologies for context defini-
tion, which are defined as the most general category of ontologies applicable across large
sets of domains. General upper ontologies are DOLCE (Descriptive Ontology for Linguistic
and Cognitive Engineering) [61], SUMO [115], Cyc [87], and SOUPA [26]. Nonetheless,
ontologies still have their drawbacks, especially in issues related to scalability and domain
specific reasoning.
Due to scalability and domain-related issues other researchers have proposed their own
ontologies like Lopez et al. [89], who proposed a generic ontology Emotions Ontology.
The ontology differentiates the concepts of the physical world and the mental world.
By relating these two worlds they achieved a model for describing emotions and their
detection and expression in systems that work with contextual and multimodal elements.
To complete their specification they used DOLCE because it defined concepts as event,
process, or action, used to contextualize other concepts in the Emotions Ontology; and
FrameNet [139] because it is better suited for modeling context as situations. Although
this work resembles our research, they did not deal with concepts as goals, likings of the
character or relation with other agents.
Nakasone et al. [112] presented a generic storytelling ontology model, the Concept
Ontology. Their idea was to define a set of topics in which the story, or part of the story,
is based, and to link them through a pseudo-temporal relation that ensures a smooth
transition between them. As this ontology was for storytelling, the classes defined were
related to scenes, acts, relations, agents participating in the story and their roles. Its
main advantage is that it provided the elements for creating a story according to narrative
principles.
Figa et al. [56] used ontologies and other components to develop an architecture for
Virtual Interactive Storytelling Agents (VISTAs) capable of interacting with the user
through natural language. First, they generated AIML scripts from live audio, video
recordings of storytellers, or from online chats, which were analyzed using the WordNet
lexical database [55] enhanced with Prolog inference rules. Then, they used ontologies
to select a storys subset expressing the focus of interest of the user. These agents were
oriented to online teaching and shared virtual environments to support learning.
Swartjes et al. [142] implemented a multi-agent framework that generated stories based
on emergent narrative. Story generation was performed in two phases: simulation and
57
CHAPTER 5. CONTEXT REPRESENTATION
presentation. During simulation, the one that concern us the most, they used an ontology
to define the settings of the story world. They used the example of a pirates world, where
the ontology defined concepts as Sword or Pirate, and relations as hate or owns.
Ontologies can also be used for the representation of the emotional output generated
by context. For instance, Obrenovic et al. [116] provided flexible definitions of emotional
cues at different levels of abstraction. Their Emotional Cues Ontology provided a language
to share and reuse information about emotions. The concepts handled by this ontology
were emotions, emotional cues (i.e. facial expressions, gestures, speech and voice), and
media (where the emotional cues are represented). Their main objective was to generate
or recognize emotional cues according to the media and the emotion felt. They did not
focus on the process of simulation of the environment that elicits the affective information.
Heckmann et al. [71] introduced the general user model ontology GUMO as part of a
framework for ubiquitous user modeling. What is interesting about this work is the model-
ing of Basic User Dimensions as: personality, characteristics, emotional state, physiological
state, mental State and facial Expression. The principal use of this data is in assessing
the actual state of the user. Regarding facial expressions, if a user shows some expression,
they represented this information in their UserJournal without interpreting the current
emotional state of the user. This work is very similar to our work but from the Ubiquitous
Computing field. Nevertheless, their inference process is related to what will occur in the
environment (ambient intelligence) and not to what will be the affective state of the user.
Chang et al. [25] presented a three-layered framework for scenario construction, formed
by: a mind model, a concept model and a reality model. The concept model provided
characters with ontology-based environmental information, so they can use ontological
inference to associate objects in the world with their goals and build plans according to
the world where they are. Nonetheless, the concept layer did not take into consideration
affective elements, which were in the mind layer.
Benta et al. [18] proposed an application that combined affect, context, ontologies and
ubiquitous computing. They presented an ontology-based representation of affective states
for context aware applications. Using the Activation-Evaluation space proposed by Cowie
et al. [32], they extended the ontology SOCAM (CONON) [149] so it can define different
States: Affective, Mental and Physiological. CONON is an ontology for modeling context
in pervasive computing environments.
Gutierrez et al. [70] developed an ontology to provide a semantic layer for concepts
related to: human body modeling and analysis, animation of virtual humans, and interac-
58
5.2. SEMANTIC MODEL
tion of virtual humans with virtual objects. This ontology was more focused on defining a
character from a physical and behavioral point of view, which can be of great help when
deciding which actions to take depending on the context that surrounds the character.
Our intention with this overview of previous works was to show how context has been
represented before, and how ontologies have also been used to represent behavior in virtual
characters. Nevertheless, we found that former research lacks of a structured definition
of the psychological characteristics that build a virtual character from an internal point
of view. Moreover, we think it is necessary to describe the process between situations
produced in certain context and the elicitation of emotions based on a belief-desire-intention
perspective. This process of event appraisal carried on by the character is what we have
intended to define using ontologies.
59
CHAPTER 5. CONTEXT REPRESENTATION
The development of ontologies usually starts by defining its domain and scope. It means
that given a motivating scenario, we should be able to answer certain questions related to
it. These questions are known as competency questions (CQs). Informal CQs are queries
which require the objects or constraints defined with the object. Fomal CQs are defined as
an entailment or consistency problem with respect to the axioms in the ontology [69]. We
have focused only on informal CQs to evaluate the expressiveness of our ontologies: Outer
World CQs and Inner World CQs.
The Outer World Competency Questions are:
1. In a certain situation (e.g. a fragment of a story, an episode of daily life), which are
the occurring events?
2. How is an event described? Which properties does it have? Which is the main action
that gives meaning to that event?
4. Where is the event happening? Can it happen in more than one location?
5. When is the event happening? How can we describe the duration of events? How
can we simulate the duration of an event?
6. Can events be organized temporarily (event 1 occurs first, then event 2, and so on)?
Can they be organized automatically, so they produce some temporal lineal story?
7. Who is performing the event? Or who is affected by the event? Are there other
agents involved? How are they related?
1. Which are the goals of an agent? What are the consequences of achieving a goal?
2. What does the agent like and dislike? Can they be physical objects or intangible
things?
4. Which emotions can the agent feel? How are they elicited?
5. What is the personality of the agent? How does it influence its emotional state?
60
5.3. EVENT ONTOLOGY
From the former analysis we decided to design and implement two different ontologies,
represented in Figure 5.2. The personalityEmotion Ontology considers all the concepts
that define a character (or agent) from a psychological-affective point of view - goals,
preferences, social admiration with other agents, and personality (inner world). The event
Ontology describes the environment that surrounds the character (outer world) based on
the events occurring in it. The following sections describe both ontologies in detail
An Action is represented with a verb. So, as in grammar a verb has complements (direct or
indirect objects) that give a meaning or complete the sentence, an Action has complements
represented by the classes AbstractEntity and PhysicalEntity.
AbstractEntity includes all intangible concepts as ideas, thoughts, dreams, or standards
(the notion of how to be socially correct).
PhysicalEntity defines all things that we can see and touch through the subclasses
SpatialLocation, MaterialThing and Agent. A PhysicalEntity has a Dimension formed by
the subclasses: Width, Height, and/or Depth. For instance, a building can be very wide
and tall, a pool can be 3 meters deep, or a floor-lamp can have a height of 1.2 meters.
61
CHAPTER 5. CONTEXT REPRESENTATION
The relation that links the concepts Event and Action are what and onEvent. The
latter specifies that an action occurs in certain event, making the relation between these
concepts cyclic. For instance, the action Eat can be linked to physical entity Bread,
and it is part of certain event. Therefore, the description logic (DL) representation is
Eat := Action
(thereAreIrrealT hings(Eat, ) thereAreObjects(Eat, Bread))
what(Event1, Eat)
SpatialThing defines where the event occurs. Its subclasses are: HumanStructure (e.g.,
buildings), NaturalStructure (e.g., beach) and OpenStructure (e.g., street).
This concept has a characteristic represented by the class AmbientCondition. Its sub-
classes Temperature and Humidity describe how cold or hot, and how wet or dry, a place
is. From these two elements Weather can be inferred, thus defined as the state of the
atmosphere with respect to temperature, and wetness or dryness.
62
5.3. EVENT ONTOLOGY
The relation that relates the event with its location is named where. For example, an
event that occurs in a Farm in sunny day can be described as
F arm := HumanStructure
hasSuperClass(F arm, SpatialT hing)
hasT emperature(F arm, 25)
hasW eather(F arm, Sunny)
An agent, represented by the class Agent, is the person or animal around which the event
happens. Therefore, it extends the subclasses Character and Animal to differentiate be-
tween human-like agents and animal-like agents.
To know exactly the role of the agent in an event, we have created the class AgentRole
with its subclasses: Execute and Receive. The subclass Execute instances those agents that
perform the action of the event (executers), while Receive instances the agents affected by
that action (receivers).
However, to specify even more the role we extended these two classes to represent the
participation of the different agents in the event. Thus, Execute presents the subclasses
ByMe and ByOther, and Receive extends the subclasses OnMe and OnOther.
- The role ByMe is taken by the agent who analyses the event and at the same time
executes the action.
- The role ByOther is taken by that agent who executes the action, but is not the one who
analyses the event.
- The role OnMe is taken by the agent who analyses the event and at the same time receives
the effects of the action.
- The role OnOther is taken by the agent that receives the action of the event, but this
agent is not the one who analyses the event.
Let us take the situation Rose buys flowers for Charlie. If Charlie is the one who
evaluates the event, the Rose has the role ByOther because she is executing the action
buy. But if Rose evaluates the event, her role would be ByMe and Charlies role
would be OnOther because he is the passive subject.
The relation that relates the event with the agent, and specifically, with its role is who.
For example, using the previous event we can describe it as
63
CHAPTER 5. CONTEXT REPRESENTATION
Charlie := Character
hasSuperClass(Charlie, Agent)
hasRole(Charlie, Charlie OnM e)
hasSuperClass(Charlie OnM e, Receive)
morning := Interval
hasSuperClass(morning, T emporalEntity)
hasBeginning(morning, time0600)
hasEnd(morning, time1200)
64
5.4. PERSONALITYEMOTION ONTOLOGY
How someone feels is the result of a process of context appraisal, which is performed using
information like ones goals, preferences, degree of admiration for others, and personality
traits.
To implement these concepts, explained in Section 4, we designed the PersonalityEmo-
tion Ontology, presented in Figure 5.4. It is worth noting that from now on we use the
term Character instead of Agent because we are considering the domain from a human
perspective. But as a Character is an Agent, then both terms can also be used indis-
tinguishable. The following subsections explain the diagram according to the main classes
pointed out by letters.
65
CHAPTER 5. CONTEXT REPRESENTATION
The class EventRelation defines the relation established between an Event, a Character
and the level of satisfaction produced in that character specified by the class EventSat-
isfactionScale. The subclasses of EventSatisfactionScale are: SATISFACTORY (S), IN-
DIFFERENT (IA), and NOT SATISFACTORY (NS), and each one has instances which
values range in a degree (0, 1).
For instance, in the event: event001:Rose goes to the movies every Saturday, we can
deduce that the character Rose enjoys going to watch films every week, and we can describe
it as
The Goals of a character are events with certain desirability degree. This degree indicates
how much a character wants that event to happen. If desirabilitydegree 0.7, then we
consider the event as a goal. The intensity of the emotions elicited due to the occurrence,
or achievement of a goal are determined by the desirability degree.
Preferences are set for physical or abstract entities. The class PreferenceRelation defines
the relation established between a Character, its Preferences and the emotional attachment
to them, expressed by the class EmotiveScale (Fig. 5.4:D).
The class Preferences groups all the instances independently of the character.
The class EmotiveScale has the subclasses: STRONGLYGOOD (SG), GOOD (G),
INDIFFERENT (IP), BAD (B), and STRONGLYBAD (SB), and each one determines the
emotion to be triggered: Love, Liking, No emotions, Disliking, Hate/Fear, respectively
(Fig. 5.4:E). The intensity of each emotion is determined by the degree of each instance
of these subclasses, where degree (0, 1). These emotions are taken from the OCC model
set based on attitudes or tastes [27].
66
5.5. EMOTION ELICITATION
PreferenceRelation allows us to model the idea of the OCC model, where attitudes of
the character towards aspects of objects lead to Like them or Dislike them; and when
there is attraction to those objects, to emotions of Love or Hate. A set of logic rules helps
to discern between fear and hate.
For instance, the event rcomedy001: Rose enjoys watching romantic comedies can
be formally described as
67
CHAPTER 5. CONTEXT REPRESENTATION
Figure 5.5 is an expansion of the red-dotted segment of Figure 5.4:F,G. It shows the
implementation of the different roles as individuals and how they are related with Emotions
and AgentAdmirations individuals. An individual is an instance of an ontology class.
68
5.5. EMOTION ELICITATION
The subclass onMe indicates that the agent receives the effects of the event. According to
the OCC theory, the effects or consequences for the self can be relevant or irrelevant. If
irrelevant, elicited emotions are Joy or Sadness.
If consequences are relevant, then there are two individuals related to Hope or Fear.
So, to know which emotions to elicit, we need to follow these rules:
- If the event is hoped and it occurs (meaning it is confirmed), the emotion Satisfac-
tion arises; but if it is disconfirmed, the emotion Disappointment arises.
- If the event is feared and confirmed, the emotion Fear-Confirmed is elicited; but if
it is disconfirmed, the emotion Relief is elicited.
We have represented these consequences as the subclasses irrelevant and relevant. ir-
relevant has the individual irrelevant emotions (Fig. 5.5: ), which is linked to the indi-
viduals Sadness and Joy through the relation hasSelfEmotion (Fig. 5.5: G,E). relevant
has two instances or individuals: onMe fear and onMe hope. The first one relates the
agent with role onMe with the emotions Fear and Relief through the relation hasSelfEmo-
tion (Fig. 5.5: ,G,E). The second one relates the agent affected by the hoped event to
Satisfaction and Disappointment through the same relation hasSelfEmotion.
For example, in the case Rose has the role Rose OnMe, first we need to see if the event
in which she participates is relevant or irrelevant for her. If the event is a Goal, then it is
relevant, otherwise, it is irrelevant.
1. If relevant, then Rose OnMe should be equivalent to one of the individuals of the
class relevant according to the following cases:
If the goal is hoped, then Rose OnMe is associated with the individual
onMe hope, and thus associated with Satisfaction and Disappointment.
If the goal is goal is feared, which means that the sentence describing the goal
has an opposite semantic meaning to the goal (e.g. a goal would be winning
the lottery, but the event is Not winning the lottery), then Rose OnMe is
associated with the individual onMe fear, and thus associated with Fear and
Relief.
The next step is to discern which emotion will be elicited. Then, we take a look at
the satisfaction level of the goal-event.
69
CHAPTER 5. CONTEXT REPRESENTATION
If Rose OnMe is associated with onMe hope and the event is Satisfactory, the
elicited emotion is Satisfaction; if the event is NotSatisfactory, the elicited emo-
tion is Disappointment.
If Rose OnMe is associated with onMe fear and the event is Satisfactory, the
elicited emotion is Relief; if the event is NotSatisfactory, then Fear is elicited.
According to the OCC theory, the action of the agent focused on the self can produce
Pride or Shame; and depending on the consequences of the event, produces Gratification
or Remorse. In our model, this is translated as:
If the event is Satisfactory, elicited emotions are positive: Pride and Gratification
If the event is NotSatisfactory, elicited emotions are negative: Shame and Remorse.
The subclass byMe represents the agents that perform the action of the event, and the
individuals are also related to Emotion through the relation hasSelfEmotion (Fig. 5.5: ).
Therefore, the individual ByMe Emotions is linked with the individuals Remorse, Shame,
Gratification and Pride.
For instance, if the individual Rose is the one performing an event, her role Rose ByMe
will be equivalent to the individual ByMe Emotions. Then, we need to see if the event in
which she participates is Satisfactory or NotSatisfactory, and the elicited emotions will be
the ones corresponding to the individuals previously enumerated.
The OCC theory proposes the set of emotions that should be elicited regarding the conse-
quences or effects that an event has on other agents. If the event is desirable for others,
elicited emotions are Happy-for or Resentment, and if the event is undesirable for others,
elicited emotions are Gloating or Pity.
70
5.5. EMOTION ELICITATION
Rose onOther is equivalent to Desirable if the consequence of the event is desirable for
her, which means that the event is Satisfactory. To discern between elicited emotions,
we consider the AgentAdmiration degree:
1. If the main agent has a Positive admiration for Rose, the elicited emotion in
the main agent is positive, Joy
2. If the main agent has a Negative admiration for Rose, the elicited emotion in
the main agent is negative, Resentment
1. If the main agent has a Positive admiration for Rose, the elicited emotion in the
main agent is negative: Pity (the main agent feels sorry for the non-satisfactory
event on Rose)
2. If the main agent has a Negative admiration for Rose, the elicited emotion in
the main agent is positive: Gloating
The OCC theory that states that the action of other agent can produce Admiration or
Reproach, and depending on the consequences of the event, it can produce Gratitude or
Anger towards that other agent.
71
CHAPTER 5. CONTEXT REPRESENTATION
In our model, the individual ByOther Emotions of the class byOther is linked to the
individual AgentAdmiration byOther of the class AgentAdmiration through the relation
hasAdmiration. In turn AgentAdmiration byOther is linked with the individuals Admira-
tion, Reproach, Anger and Gratitude (Fig. 5.5: , G).
Taking as example Roses role Rose byOther, which is equivalent to AgentAdmira-
tion byOther:
If the event is Satisfactory and the main agent has a Positive admiration for Rose,
the elicited emotion is Admiration
If the event is Satisfactory and the main agent has a Negative admiration for Rose,
the elicited emotion is Gratitude
If the event is NotSatisfactory and the main agent has a Positive admiration for
Rose, the elicited emotion is Reproach
If the event is NotSatisfactory and the main agent has a Negative admiration for
Rose, the elicited emotion is Anger
Figure 5.6 shows a diagram with all the elicited emotions according to the type of event
(Positive (P) or Negative (N)) and the type of agent admiration (Positive (+) or Negative
(-)), which summarizes what was explained above.
72
5.6. GUIDELINE TO USE THE ONTOLOGIES
As seen in the previous sections, the Event ontology provides all the elements to represent
an event, while the PersonalityEmotion ontology provides the relations between the event
and the character, as well as all elements related to the characters internal aspects (goals,
preferences, admiration for other agents, personality) used in event appraisal.
The following is a guideline for representing a situation in certain scenario:
1. Situation Identification. First of all we need to describe what the characters will
go through as if it were a story. In this way we can start identifying all the elements
of the ontologies. For example,
Rose started her day at 6 a.m. this morning. While she was having breakfast, she
heard about train strike in the radio. This was very inconvenient, because she had an
early appointment with Charlie at the office and with the strike she wont get there.
Each event is seen as a predicate that includes a subject (agent), verb (action) and
predicate (ambient conditions, physical and abstract things - objects, other agents,
location). Therefore, once we define the instances of each class in the ontology, we
assign a level of Satisfaction to each event regarding the agent that appraises it.
For instance, from the situation Rose started her day at 6 a.m. this morning, we
can get the Event instance:
3. Character Preferences and Goals definition. Once all elements that form the
event are identified, we assign values to them using the PersonalityEvent Ontology
(Fig. 5.4:B,C). The steps to follow are:
73
CHAPTER 5. CONTEXT REPRESENTATION
4. Agent Admiration for other agents. To define the admiration that the main
character (the one from whose point of view the event is evaluated) feels for other
agents, we need to define the instances of AgentAdmiration in the story. In the former
example the other agent is Charlie, so the event can be defined as:
event004 := Rose attends to an appointment with Charlie
First of all, we need to define the roles of Rose and Charlie to proceed with the
identification of the corresponding instances. As this event might be ambiguous, we
have two options:
(1) Roses role is Rose byMe and Charliess role is Charlie onOther
(2) Roses role is Rose onMe and Charliess role is Charlie byOther
In case (1), we would need to define the event as desirable or undesirable for Charlie.
As this event is neither one nor the other for Charlie, then we try with case (2).
Thus, following Fig. 5.5: :
74
5.6. GUIDELINE TO USE THE ONTOLOGIES
5. Personality definition. Using the FFM the personality traits for each character
are set. For instance, we can define Roses personality as:
extraversion rose := Personality
personalityDegree(extraversion rose, 0.2)
agreeableness rose := Personality
personalityDegree(agreeableness rose, 0.3)
conscientiousness rose := Personality
personalityDegree(conscientiousness rose, 0.9)
openness rose := Personality
personalityDegree(openness rose, 0.1)
neuroticism rose := Personality
personalityDegree(neuroticism rose, 0.95)
75
CHAPTER 5. CONTEXT REPRESENTATION
8. Contained Events (Optional). In case there are contained events (given by the
relationship contains), the appraisal is done in the main event following the steps
described above, but taking into consideration the next issues:
5.7 Implementation
To implement the proposed Semantic Model, we divided the tasks in two main parts:
2. Implementation of the interface: using JDK 1.6.0 12 and JENA libraries for ontology
support.
In order to implement our ontologies, we needed to consider various aspects like the lan-
guage to use, how to create inference rules, or which other ontologies could be useful to
complete the knowledge we wanted to represent.
The answers to these issues are explained in detail in the following subsections.
76
5.7. IMPLEMENTATION
The W3Cs Web Ontology Language (OWL) [68] is considered by many as the prospective
standard for creating ontologies on the Semantic Web.
OWL features two core elements: classes and properties. A class represents a set of
individuals, all of which also belongs to any superclass of the class. A property specifies
the relation between an instance of the domain class to an instance of the target class [25].
Besides that, OWL provides individuals and data values that are stored as Semantic Web
documents.
OWL was designed to be used by applications that need to process the content of
information instead of just showing it to the user. OWL provides a better machine inter-
pretability of Web content than that offered by XML, RDF, and RDF Schema (RDF-S)
because it provides the formal semantics plus additional vocabulary [100].
To model the ontologies, we use OWL because it is the standard language and because
of its rich expressiveness and solid foundation for concept modeling and knowledge-based
inference. The ontologies editor we have worked with is Protege version 4.0, a free and open
source ontology editor and knowledge-base framework [57]. Figures 5.7, 5.8, 5.9, 5.10, 5.11
and 5.12 show the ontology diagrams for the different concepts of the ontologies defined
using OWL and Protege.
77
CHAPTER 5. CONTEXT REPRESENTATION
78
5.7. IMPLEMENTATION
79
CHAPTER 5. CONTEXT REPRESENTATION
80
5.7. IMPLEMENTATION
External Ontologies
Many times the ontologies proposed for certain domain are not enough, and one can found
other ontologies already defined by other researchers that help to describe all the required
knowledge.
In our case, we have used the WordNet [107] library for ontology Java processing,
named RiTa.WordNet [75]. It defines the verbs as instances of the class Action, nouns
as instances of the class Preferences, and adverbs to give semantic meaning to the instances
of Event.
RiTa.WordNet is the API that provides simple access to the WordNet ontology. It is
an easy-to-use natural language library that provides simple tools for experimenting with
generative literature.
The other external ontology we have used is the OWL-Time ontology, which is an
ontology of temporal concepts. It provides a vocabulary for expressing facts about topo-
logical relations among instants and intervals, together with information about durations,
and about datetime information [73].
We have used the OWL-Time ontology to specify the instances of TemporalEntity,
defined in our Event Ontology. Thus, we could represent the temporal characteristics of
an event: when it started, how long it lasted (minutes, hours), and so on.
Inference Rules
Inference rules are useful to define particular scenarios, which can be obtained from data
previously defined. The following are the scenarios where rules are used:
In our model, most of the rules allow the elicitation of the emotion, or set of emotions,
depending on the level of satisfaction of the event (EventSatisfactionScale) and role of
the character (AgentRole). For instance, if we have an agent with the role byMe (e.g.
Rose ByMe), it can be associated to the already defined individual ByMe emotions,
and with the rule, emotions linked to this individual will be triggered.
Rules are also used when the level of admiration (AgentAdmiration) of one agent for
another needs to be taken into account. In the same way as with the agent role,
an individual of a subclass of AgentAdmiration (e.g. negative aa, which indicates a
Negative admiration) can be associated through inference rules with a certain agent
81
CHAPTER 5. CONTEXT REPRESENTATION
role (e.g. Rose ByMe), and therefore, only the corresponding negative emotions will
be elicited.
Finally, rules can be used to enhance the personality traits levels of the character,
according to the rules defined by Poznanski and Thagard [126]. They imposed limits
on how much personality can change due to the environment. Once this limit is
reached, the environment does not change the personality anymore. They considered
five types of situations: friendly, hostile, explore, persist and stressful. However, as
we have just two relevant types of events, or situations, we will consider the rules
just for friendly (SATISFACTORY) and hostile/stressful (NOTSATISFACTORY).
Table 5.1 presents the rules proposed by them, and adapted to our work.
Appendix A shows all the rules defined in Java using Jena libraries, which are triggered
at the beginning of the application execution.
Jastor
Jastor is a open source Java code generator that emits Java Beans from Web Ontologies
(OWL), and generates Java interfaces, implementations, factories, and listeners based on
the properties and class hierarchies in the Web Ontologies [143].
82
5.7. IMPLEMENTATION
The advantage of using Jastor is that it converts all the classes defined in the ontologies
to their equivalent Java classes, including individuals already defined in Protege. If one
needs to add more individuals, it just requires to modify the .java files. Nevertheless, to
rewrite the .owl files, it is necessary to do it through the data stored in a database, which is
done using JENA. Figure 5.13 shows an extract of the implementation of the class ByOther,
which was automatically generated by Jastor.
Once all the classes corresponding to the ontology entities were generated, these were
used and modified by the application through functions of the JENA library.
With Java JDK 1.6.0 12 and JENA library [72] we were able to perform operations on the
ontologies, as well as inference that was done using JENA rules. JENA is a Java framework
for building Semantic Web applications. It provides a programmatic environment for RDF,
RDFS and OWL, SPARQL and includes a rule-based inference engine (JENA rules).
One of the advantages of JENA is that being developed in Java, it is applicable to
various environments. In addition, it is open source and strongly backed up by solid
documentation. Also, regardless to the schema and the used data models, JENA can
simultaneously work with multiple ontologies from different sources, and it can be bound
to SQL databases from different vendors.
Another powerful feature of the JENA framework is the inference API. It contains
several reasoner types, which efficiently conclude new relations in the knowledge graph.
Among the reasoners there are: RDF(S), OWL, Transitive and Generic reasoners. JENA
also works with third party reasoners as Pellet.
83
CHAPTER 5. CONTEXT REPRESENTATION
Interface
To create the GUI that allowed us to manipulate the ontology classes, we used JFC/Swing.
The idea for creating an interface like this is that any user can create characters and
define for each of them the events, goals, preferences and admiration for other agents,
which will define their context. Then, all these data would be stored in a database (using
JENA API), so it can be manipulated and reused to create new scenarios, or similar ones
with different emotional outputs.
In this way, just by assigning values for the concepts of the ontologies, the application
is capable of inferring and creating the relations between those concepts, giving as a result
a set of emotions which are felt by the character. As a final goal, a database of contextual
situations could be constructed, containing any kind of related data to be reused by the
defined ontologies. Figure 5.14 shows two windows of the GUI which correspond to the
main window and the physical entity window.
84
5.8. USE CASE
Event Satisfaction: The contained event is evaluated from the perspective of the
character that performs the main event: Blanche. She considers this event as NOT-
85
CHAPTER 5. CONTEXT REPRESENTATION
Admiration Degree: If Blanche is the performer of the main event then her role is
ByMe, the role of Jane who performs the contained event is ByOther. Admiration
of Blanche for Jane is set to NEGATIVE = 0.6.
Personality: Blanche is a little bit extroverted (0.4), extremely agreeable (0.99), mod-
erately conscientious (0.8), not neurotic (0.2), and extremely open (0.99). On the
other hand, Jane is extremely extroverted (0.99), disagreeable (0.2), not conscientious
(0.2), extremely neurotic (0.99), and somewhat open (0.6).
Emotion Elicitation: Using Figure 5.6, we first elicit emotions using the contained-
characters role (Jane, ByOther) and the contained-event satisfaction (NOT SATIS-
FACTORY = 0.8), from the main-characters perspective (Blanche). Elicited emotion
is anger = 0.6.
Now we need to evaluate the event using the main-characters role and contained-
event satisfaction. It means, the role of Blanche (ByMe) and event Satisfaction is
NOT SATISFACTORY = 0.8. Elicited emotion are shame = 0.8 and remorse = 0.8.
To consider the main event satisfaction, we need to use a logic rule to decide for it.
The rule is as follows:
IF contained-event.SatisfactionScale = NOT SATISFACTORY
AND action = NOT believe
THEN main-event.SatisfactionScale = SATISFACTORY AND byMe.emotions =
pride = 0.8 OR gratification = 0.8
It means that as Blanche does not think that a Not Satisfactory event will occur,
then it becomes Satisfactory. As we can see, elicited emotions for role ByMe are
different; therefore, we select the ones obtained from the rule.
The final set of emotions is: Anger = 0.6, Pride = 0.64, Gratification = 0.64 . From
this set we chose the ones with the greatest values.
5.9 Summary
The novelty of our approach is the representation of context by means of a semantic model,
not only as events in the world, but also as the internal characteristics of the character which
86
5.9. SUMMARY
when related with the events, give as result believable emotional responses. This chapter
presented an overview of previous works on context representation and how they have been
related with the creation of virtual characters; as well as the description, definition and
implementation of the ontologies that form part of the semantic model. Finally, an use
case demonstrated not only how to use these ontologies, but the obtained results after the
simulation of a set of events.
87
CHAPTER 5. CONTEXT REPRESENTATION
88
Chapter 6
It is known that behavior, social interaction and communication, action planning and
response to the environment are guided and regulated through emotions [17]. Neverthe-
less, other traits like mood and personality have also a strong influence on emotions and
behaviors.
Hence the achievement of more complex and richer affective characters, or agents, will
depend on the interrelation between emotions, mood and personality.
In the previous chapter we went through the process of context appraisal to obtain a set
of emotions that are experimented by the character in certain situation. In this chapter,
we explain how to enrich those emotions with mood and personality. As result we have
a set of pleasure, arousal and dominance values, corresponding to the resultant mood of
the character.
This chapter is organized as follows. Section 6.1 introduces the affective layer used
for regulation and generation of moods, based on the PAD model proposed by Albert
Mehrabian [105]. Section 6.2.1 provides a detailed description of the affective model used
to process the emotions received from the semantic layer (Chapter 5) in order to generate
the values for different moods. Finally, Section 6.3 offers a summary of the chapter.
89
CHAPTER 6. AFFECTIVE MODEL FOR MOOD GENERATION
In our work emotions are generated based on the OCC model and personality traits
are chosen from the Five Factor Model (FFM). To relate emotions, moods and personality
in the same space, we use the Pleasure-Arousal-Dominance space (PAD) because it allows
us to obtain mood values in terms of pleasure, arousal and dominance dimensions.
The PAD space is considered a dimensional model, and the motivation to choose it for
mood representation is derived from the ideas explained by Cochrane [29]. One attraction
for a dimensional approach is that it respects a fundamental observation that emotions can
vary very smoothly, and the information they provide is typically dynamic in content. The
other idea relates to language; therefore extracting dimensions from language can allow one
to transcend its constraints to some extent and break down affective terms into descriptive
terms that convey the same meaning for different people both within and across cultures.
90
6.1. AFFECTIVE MODEL
The notations +P and -P for pleasure and displeasure, +A and -A for arousal and
nonarousal, and +D and -D for dominance and submissiveness, respectively, are used
throughout this work.
Pleasuredispleasure distinguishes positive affective states from negative ones. Arousal-
nonarousal is defined in terms of a combination of mental alertness and physical activity.
Dominance-submissiveness is defined in terms of control versus lack of control over events,
ones surroundings, or other people. The resulting octants and mood categories, corre-
sponding to various combinations of high versus low pleasure, arousal, and dominance [104].
These are indicated in Table 6.1.
Exuberant (+P +A +D) Bored (-P -A -D)
Docile (+P -A -D) Hostile (-P +A +D)
Dependent (+P +A -D) Disdainful (-P -A +D)
Relaxed (+P -A +D) Anxious (-P +A -D)
91
CHAPTER 6. AFFECTIVE MODEL FOR MOOD GENERATION
for choosing this model. For example, one could be able to map the Big Five per-
sonality model, or de AB5C personality model, into the PAD space and obtain the
same set of temperaments, or moods [104].
Widely used. In the last few years this model has been one of the most used to rep-
resent personality, emotions and mood. According to Garvin [62], referencing other
authors, the components within the PAD scale have shown to have good reliability
and nomological validity through the many studies that have utilized them. Actually,
the PAD space has not only been used by researchers in Affective Computing like
Gebhard [63], Becker-Asano [12], Kessler et al. [82], Kasap et al. [80], Ben Moussa and
Thalmann [111], Courgeon et al. [30] among others, but also by researchers in areas
like computer vision to recognize facial expressions [24], in design of theme parks [108],
for integration of emotional strategies in spoken dialogues [19], or for marketing and
communications, product development and personnel management [103].
Easiness of assessment. To evaluate the affective output of the PAD space there
is a non-verbal, visually oriented questionnaire based on the three dimensions of
pleasure, arousal and dominance: the Self-Assessment Manikin (SAM) [22]. SAM
depicts each PAD dimension with a graphic character arranged on a linear scale, and
subjects rate each input according to this character, obtaining fast evaluations and
very accurate results.
For the aforementioned reasons, the PAD model would provide us the theoretical frame-
work to create more believable characters capable of react in different situations. The
following section explains how emotions elicited in the Semantic Layer are mapped into the
PAD Space and through the Affective Layer, moods are generated and expressed in terms
of pleasure, arousal and dominance.
92
6.2. AFFECTIVE LAYER
A mood M is defined as a point in one of the eight octants of the PAD space. Therefore,
it is defined by the values of pleasure, arousal, and dominance as M = (P, A, D), where
P [1, 1], A [1, 1], D [1, 1]. Figure 6.3 shows the mood in the space.
The mood intensity m is defined by its distance to
the zero point ofthe PAD mood
~
space, and it is computed as the norm of the vector OM denoted by P 2 + A2 + D2 .
Because this three dimensional space has maximum absolute values of 1.0, the longest
distance in a mood octant is 3.
As mentioned by Gebhard in his research [64], people have better understanding of
descriptions when quantified words are used instead of numbers. Hence we applied the
same categorization for mood intensities:
93
CHAPTER 6. AFFECTIVE MODEL FOR MOOD GENERATION
slightly, if m [0.0, 31 3]
moderate, if m ( 31 3, 32 3]
highly, if m ( 32 3, 3]
Personality can also be mapped into the PAD space as the ground, or default mood,
G, which is the mood start value. It is considered the normal state of the character
according to its personality. This concept is based on the relationship established by
Mehrabian between the Five Factor Model and the PAD traits, expressed in Eq. 6.1.
For example, Eq. 6.2 computes the default mood of a person with a very extro-
verted and very agreeable personality (personality = (0.1, 0.2, 0.9, 0.9, 0.1)). Then,
G = (0.739, 0.2280, 0.311) which is translated into a moderate Exuberant mood.
94
6.2. AFFECTIVE LAYER
Finally, the emotions felt by the character at certain moment are also mapped into the
PAD space. These emotions are the ones obtained from the semantic model and influence
the change of mood in the character.
An emotion E is located in the space according to its values of pleasure, arousal and
dominance, which were obtained empirically by Mehrabian and completed by Gebhard [63,
65]. Table 6.2 shows these values, which indicate how much of pleasure, arousal and
dominance exists in each emotion. The intensity e of an emotion is given by the values
obtained in the semantic model. Figure 6.4 shows how the default mood and emotions are
represented.
Emotion P A D Mood
Admiration 0.5 0.3 -0.2 +P+A-D Dependent
Anger -0.51 0.59 0.25 -P+A+D Hostile
Disliking -0.4 -0.2 0.1 -P-A+D Disdainful
Disappointment -0.3 -0.4 -0.4 -P-A-D Bored
Sadness -0.5 -0.42 -0.23 -P-A-D Bored
Fear -0.64 0.60 -0.43 -P+A-D Anxious
Gloating 0.3 -0.3 -0.1 +P-A-D Docile
Gratification 0.6 0.5 0.4 +P+A+D Exuberant
Gratitude 0.4 0.2 -0.3 +P+A-D Dependent
Hate -0.6 0.6 0.3 -P+A+D Hostile
Hope 0.2 0.2 -0.1 +P+A-D Dependent
Joy 0.4 0.2 0.1 +P+A+D Exuberant
Liking 0.40 0.16 -0.24 +P+A-D Dependent
Love 0.3 0.1 0.2 +P+A+D Exuberant
Pity -0.4 -0.2 -0.5 -P-A-D Bored
Pride 0.52 0.22 0.61 +P+A+D Exuberant
Relief 0.2 -0.3 0.4 +P-A+D Relaxed
Resentment -0.3 0.1 -0.6 -P+A-D Anxious
Reproach -0.3 -0.1 0.4 -P-A+D Disdainful
Remorse -0.2 -0.3 -0.2 -P-A-D Bored
Satisfaction 0.50 0.42 0.47 +P-A+D Exuberant
Shame -0.3 0.1 -0.6 -P+A-D Anxious
When more than one emotion are active at the same time, it results in different points
located in the same, or different octants. Therefore, we need to compute another point
that represents all of them, which will induce the displacement of the mood. This point
95
CHAPTER 6. AFFECTIVE MODEL FOR MOOD GENERATION
is named emotional center E, and it is the center of mass of all the emotions Ei . Eq. 6.3
depicts the intensity of the emotional center, computed as the norm of this vector from the
origin of the PAD space. Figure 6.5 shows how the emotional center is represented.
Pn
Ei ,AEi ,DEi )ei
i=1 (PP
E = n , where ei (0.0, 1.0] (6.3)
i=1 ei
96
6.2. AFFECTIVE LAYER
Current Mood
When the default mood G is changed by the emotional center E in an instant t, it becomes
the current mood of the character. The current mood, denoted by M(t), is computed
through the center of mass between G and E. Equation 6.4 shows this computation,
where M(t) is the current mood, G is the default mood, g its intensity, E is the emotional
center and e its intensity. Figure 6.6 shows the current mood M(t).
g G + e E
M(t) = (6.4)
g + e
New Mood
In a more realistic model, when the character experiments new emotions, these will not be
affected by the default mood G, but by the current mood M(t).
Then, a new mood at instant t + 1, M(t + 1), is the result of the emotions experienced
at instant t (represented by the emotional center E) and the current mood M(t). Figure 6.7
shows the generation of the new mood M(t + 1). Again, the new mood is computed as the
center of mass between the current mood and the emotional center.
97
CHAPTER 6. AFFECTIVE MODEL FOR MOOD GENERATION
The new mood represents the change of mood of the character along time. It is worth
noting that these changes are progressive, as it happens for people in their daily lives. It
will not happen that the character changes its mood from one extreme of the PAD space
to the other. The reason is that, according to Watson [150], extremely high levels of one
type of mood are rare in everyday life. Instead, daily experiences show low to moderate
intensity states. For example, someone who is mildly nervous may be very enthusiastic.
Decay
Finally, an important aspect that has to be considered is decay. It occurs when the character
has no active emotions, which means that no new emotions have been elicited, and therefore
his or her current mood will tend to go back to the ground, or default mood.
This change has to be progressive, that is why we compute the decayed new mood,
M(t + 1) as the center of mass between the default mood G and the current mood M(t),
according to Eq. 6.5. By successive repetitions of this equation, the current mood will
eventually become the default mood, unless new emotions are experienced. Figure 6.8
shows how the vector M(t) is moved towards G in the PAD space.
g G + m M(t)
M(t + 1) = (6.5)
g+m
98
6.3. SUMMARY
6.3 Summary
As can be seen, our model is a modified version of the ALMA model. We have decided for
computations based on center of mass, contrary to the vectorial computations in ALMA,
because it allowed us to represent in a simplified and effective way the influence of the
points representing personality and emotions on mood.
In this way, we obtained a model that served well enough to our purposes, allowing us
to represent a characters mood and its changes along time depending on the personality.
The representation was based on terms of pleasure, arousal and dominance, providing a
novel way for mood simulation and visualization.
99
CHAPTER 6. AFFECTIVE MODEL FOR MOOD GENERATION
100
Chapter 7
No man, for any considerable period, can wear one face to himself, and another to the
multitude, without finally getting bewildered as to which may be the true.
Nathaniel Hawthorne
Non-verbal expressions of affect can be performed using voice, gestures, body positions,
and facial expressions. For example, anger can be manifested through a raise in the tone
of the voice; but if the person in anger is introverted, then he or she would just frown.
In this work, the main focus are facial expressions because they have been found to be
the richest source of information about emotions, and one of the richer and more accurate
ways of expressing affect [46].
One of the first scientists who studied the facial expressions in men and animals was
Charles Darwin. All his observations were captured in his book The Expression of Emo-
tions in Man and Animals [33]. He associated different facial movements with certain
emotions, which were observed in different persons in different countries under the same
emotional state. This led to the formulation of the principle of the universality of emotions.
In the previous chapters we went through the process of context appraisal to obtain a
set of emotions, and explained how these emotions are modulated and contribute to the
elicitation of moods. In this chapter, we explain how to visualize these affective traits.
This chapter is organized as follows. Section 7.1 gives an overview of previous works
on visualization of facial expressions. First the two most used standards for creating
parameterized facial expressions, MPEG-4 and FACS, are explained. Then, it is presented
101
CHAPTER 7. VISUALIZATION OF AFFECT IN FACES
the two facial animation engines and other applications used for generation of the facial
expressions. Afterwards, the algorithms and techniques used in the generation of emotion
and mood expressions, as well as personality cues are detailed. Finally, a summary of the
chapter is presented.
Facial expressions should represent a strong emotion felt by the character at some
instant. We consider two types of emotions: universal ones (joy, anger, sadness,
disgust, surprise, and fear ) and intermediate emotions (the ones proposed in the
OCC model, e.g. disappointment, remorse). As there are no parameters established
for the intermediate ones, these should be obtained from already defined emotions.
Emotions should have different intensities. For example: does the lowering of the
eyebrow control the intensity of an angry face, and will a greater lowering of the
eyebrows make the person look angrier? [54].
Facial expressions should express the mood in an interval of time. Regarding mood,
Faigin sets the question: to what extent are day-to-day moods visible on the face?
To answer it, he cites the example of fear: ... if we are terrified enough our response
will be etched on the face. But in the case of a long-term mood like anxiety, Faigin
says that it depends on coping as well-socialized adults, we still go about our daily
life in spite of a disruptive inner mood of anxiety.
Considering the requirements mentioned above, this section describes the last mod-
ule of our computational model, which concerns to the generation and visualization of
facial expressions in virtual characters. Figure 7.1 shows an extract of the framework that
corresponds to the visualization layer.
102
7.2. EXPRESSION CODING SYSTEMS
As can be seen, two types of facial expressions are visualized: expressions for emotions
and expressions for moods. Expressions of emotions are generated using the output of the
Context Representation module -or Semantic model. For their visualization we use the
MPEG-4 standard.
Expressions of moods are generated from the output of the Affective Model. The
obtained values are mapped into the Pleasure-Arousal-Dominance (PAD) space, which
allows to visualize any variation of mood. In this case, only FACS were used.
Concerning personality, its visualization relies on visual cues, which in this work are
head pose and eye gaze. In this case, also FACS is the only standard used.
The visualization of facial expressions is considered one of the most challenging tasks in the
field of animation. On the one hand, the slightest inconsistency can be rapidly recognized
by the human eye and lead to a lack of believability; and on the other hand, the human
face is very complex, and to capture all its details can become a very difficult task.
103
CHAPTER 7. VISUALIZATION OF AFFECT IN FACES
The techniques used for generation of facial expressions can be grouped into: geometry-
based and image-based techniques.
Geometry-based techniques include key-framing with interpolation, direct parameter-
ization, pseudo-muscle-based approach and muscle-based approach. Image-based tech-
niques could be divided into morphing between photographic images, texture manipulation,
image blending and vascular expressions [129].
There is another method called performance-driven methods, which captures real peo-
ples movements and uses them to animate virtual characters. To implement techniques
that belong this last method two standards or coding systems have been stated as the most
important ones: the MPEG-4 standard and the Facial Animation Coding System (FACS).
7.2.1 MPEG-4
MPEG-4 is an ISO/IEC standard developed by the MPEG (Moving Picture Experts
Group) [58], [144]. One of the objects provided by the standard is a facial object that
represents a human face, as well as facial data: Facial Definition Parameters (FDPs),
Facial Animation Parameters (FAPs), and Facial Animation Parameters Units (FAPUs).
Facial Definition Parameters (FDPs) are defined by the locations of feature points
and are used to customize a given face model to a particular face. There are 84 feature
points, which are used as reference to calculate the facial movement. Figure 7.2 shows
how FDPs are defined in a human face.
Facial Animation Parameter Unit (FAPUs) are the units used to defined FAP values.
A FAPU is computed from spatial distances between major facial features on the
model in its neutral state, and they are defined as fractions of distances between key
facial features (Figure 7.3). Measurement units are presented in Table 7.1.
104
7.2. EXPRESSION CODING SYSTEMS
105
CHAPTER 7. VISUALIZATION OF AFFECT IN FACES
Pasquariello and Pelachaud [122], who developed the Simple Facial Animation Engine
(SFAE) to animate Greta, a 3D facial model compliant with MPEG-4 specifications.
Courgeon et al. [31] developed MARC, an MPEG-4 based facial animation system
that also supports FACS action units based animation and BML (Behavior Markup
Language) external control.
Kasap et al. [80] had Eva, a virtual character whose facial movements for each type
of emotional response were encoded using the MPEG-4 standard.
Malatesta et al. [92] used MPEG-4 facial animation parameters to obtain facial ex-
pressions for predicted intermediate and final emotional states.
Raouzaiou et al. [131] proposed a framework based on the MPEG-4 standard [58] to
achieve two goals: modeling of primary expressions using MPEG-4s Facial Anima-
tion Parameters (FAPs); and development of a rule-based technique for analysis and
synthesis of intermediate facial expressions.
Arya et al. [6] proposed iFace, a general-purpose software framework that implements
the face multimedia object, which are based on top of the MPEG-4 standard.
Balci [8] developed XFace, a set of open source tools for creation of embodied conver-
sational agents using MPEG4 and keyframe based rendering driven by SMIL-Agent
scripting language.
The Facial Action Coding System (FACS) is a method introduced by Paul Ekman, Wallace
V. Friesen, and Joseph C. Hager to measure facial behaviors [48] and to systematically
categorize facial expressions. It was developed by determining how the contraction of each
facial muscle (singly and combined) changes the appearance of the face.
FACS measurement units are Action Units (AUs), and not muscles. One of the reasons
for this is that some changes involve more than one muscle, and then they are mapped
into one AU. Another reason is that changes produced by one muscle are separated into
different AUs to describe the actions produced by the same muscle.
106
7.2. EXPRESSION CODING SYSTEMS
FACS consists of 56 AUs in total, 44 of those used to describe facial expressions and
the remaining 12 to describe head and eye movement. Appendix B has a list with all the
existent AUs.
There are FACS that are more accentuated in one side of the face or in both. From this
point of view, FACS can be unilateral or bilateral. For example, in a smile the AUs are
present in both sides of the face. Therefore, we can read: AU12 (lip corner puller ). But in
the contempt expression, the AU12 is present just in one side of the face. Therefore, we
read it as: L/R12, meaning that it can be the left (L) or the right (R) side. Other locations
of the AUs would be bottom (B) and top (T). Nevertheless, in our work for easiness of use,
we will treat AUs as unilateral.
Regarding previous works, FACS have been mainly used in the field of facial expression
recognition. In the field of generation of facial expressions, FACS have been used for quite
some by:
Platt and Badler [124], who designed the first model based on muscular actions using
FACS as basis for the control of facial expressions.
Grammer and Oberzaucher [67] modeled all AUs as a system of morph targets at
their maximum contraction directly on the head mesh in a modeling program.
Bee et al. [14] created a facial animation system based on Ekmans FACS, which
allows the creation and animation of facial expressions.
107
CHAPTER 7. VISUALIZATION OF AFFECT IN FACES
The game engine of the University of Augsburg is a facial animation system based on FACS,
designed to create facial expressions for animated agents in an easier, more intuitive and
faster way than conventional animation tools [14].
To implement FACS, morph targets (or blend shapes) were used. In the implementation
of the 3D model Alfred, each morph target corresponds to one AU. In total, 23 AUs were
implemented.
Regarding Alfred, he was modeled using Luxology modo and Autodesk 3D Studio
Max, through subdivision surfaces and sculpting techniques. Textures are either digital
handpainted (color & specular map) or baked from a high-resolution mesh (normal map).
Head has a polycount of about 21.000 triangles.
108
7.3. FACIAL ANIMATION ENGINES AND APPLICATIONS
The created facial expressions can be saved and loaded in a XML format. The config-
uration of the controllers, the basic expressions and their reduced controller sets are based
on data from the FER Database. Figure 7.4 shows the development environment, the XML
file that is used to set up the AUs and the resultant face.
Figure 7.4: Game Engine of the University of Augsburg: development environment, con-
figuration file and facial mesh
The possibility to evaluate the influence of gaze orientation and head tilt in the
affective output, given that the engine allows head and eye movement.
The facial animation system can render the facial expressions in real time.
109
CHAPTER 7. VISUALIZATION OF AFFECT IN FACES
XfaceEd editor provides an easy to use interface to generate MPEG-4 ready meshes from
static 3D models. The XfacePlayer is a sample application that demonstrates the toolkit
in action; and the XfaceClient is used as the communication controller over network.
All the pieces in the toolkit are operating system independent, and can be compiled
with any ANSI C++ standard compliant compiler. For animation, the toolkit relies on
OpenGL and is optimized enough to achieve satisfactory frame rates (minimum 25 frames
per second are required for FAP generating tool) with high polygon count (12000 polygons).
The Xface module we use is XfaceEd, where we import the facial mesh in VRML
(Virtual Reality Modeling Language) format. Once in the editor, we can assign values to
the FDPs and FAPs parameters. In one of the tabs of the editor we define how each FDP
will influence its neighbor vertices according to muscle models (Fig. 7.5).
Figure 7.5: Tab in XfaceEd where FDPs are assigned to the facial mesh
In the FAPs previsualization tab, XfaceEd loads the facial mesh from a file with all
the FDPs definitions. Then, it loads the FAP values which are specified in a file of FAPs
frames. Finally, it decodes these values into the face and the expression can be played.
XfaceEd also offers the possibility to change and visualize each FAP at a time. Through
a modification made in the Xface editor, we are able to change all the FAPs directly into
the model and write them into the FAPs file.
110
7.4. VISUALIZATION OF EMOTIONS
What is a basic emotion and what are emotions are topics which answers still raise
debates between psychologists. In our research we use as universal emotions the ones
proposed by Ekman: Joy, Sadness, Anger, Disgust, Fear, and Surprise, mainly because
they are perceived as such in different cultures and because they are the state of art of
universal or basic emotions.
When generating expressions using MPEG-4, each emotion has a set of FAPs. Let Ei
be the i-th emotion generated and Pi the set of activated FAPs for Ei .
Each FAP j in Pi has a range of variation Xi,j [minV alue, maxV alue], indicating
the minimum and maximum values along which the FAP can be displaced. As for j, it is
defined as j = {1, ..., 64} given that the standard defines 64 FAPs.
The variation ranges Xi,j for universal emotions were obtained by manual manipulation
of FAPs to simulate the expressions of joy, sadness, fear, disgust, anger, surprise, and
neutral drawn by Faigin [54]. These models are seen in Figure 7.6, and the generated
universal facial expressions are shown in Figure 7.7.
As mentioned in previous chapters, intermediate emotions are taken from OCC model,
except for happy-for and fear-confirmed because their expressions can be represented as
joy and fear, respectively.
Based on the work of Raouzaiou et al. [131] we generated a set of intermediate emotions
using the MPEG-4 standard. They identified eight fundamental emotions: acceptance,
fear, surprise, sadness, disgust, anger, anticipation and joy, which were the starting points
for interpolation. There were two different ways to generate new expressions: if the new
emotion En is very similar to the fundamental emotion Ei , i.e., if their facial expressions
differ mainly in strength of muscle contraction, then the new expression En can be com-
puted as a category of the expression Ei . If the new emotion En does not clearly belong
111
CHAPTER 7. VISUALIZATION OF AFFECT IN FACES
Figure 7.6: Faigins universal facial expressions. Up: Anger, Disgust, Fear. Down: Joy,
Sadness, Surprise
Figure 7.7: Universal facial expressions generated using MPEG-4. Up: low intensities of
Anger, Disgust, Fear, Joy, Sadness and Surprise. Down: high intensities of Anger, Disgust,
Fear, Joy, Sadness and Surprise
112
7.4. VISUALIZATION OF EMOTIONS
Instead of combining the data from the Whissell and Plutchik [132] studies, as
Raouzaiou et al. did, we only used the data from Whissell. The reason to use only
the Whissells Dictionary of Emotions [153] is that it provides a complete list of terms
with affective connotations described in terms of their activation (or arousal) and evalua-
tion (or pleasantness). From these values, we can locate the emotions in a 2D space and
compute their angular distances, establishing similarity among them. Figure 7.9(a) shows
the emotions located into the activation-evaluation space.
Figure 7.9: Whissell Emotions: (a) Not centered values, (b) Centered values
113
CHAPTER 7. VISUALIZATION OF AFFECT IN FACES
For a better visualization of the angular distances we center the emotions with respect
to the origin (0, 0) using the mean values of the activation and evaluation dimensions
(a = 4.00 and e = 4.00) obtained by Whissell. In this way, emotions can be spread along
negative and positive axis maintaining the angular proportion between them.
Equation 7.1 shows how to obtain the angle of an emotion with its centered activation
value acenter = a a and centered evaluation value ecenter = e e. Then, we use the rules
in Table 7.2 to locate the angle in its corresponding quadrant, given that the returned
is between 0 and 90 . Figure 7.9(b) shows the centered emotions.
acenter
= arctan (7.1)
ecenter
Once emotions have been properly located in the activation-evaluation space we can
decide if an intermediate emotion will be obtained as a category or as a combination of
universal emotions.
In this work, the universal emotions used to obtain the intermediate ones, as well as
their activation, evaluation and angular values are given in Table 7.3.
In the case of categorization, Method 1 is applied. It changes the intensity of an
universal emotion, resulting in an intermediate emotion that is a category of one universal
emotion. Another way to realize if an emotion is in certain category is its closeness to
certain universal emotion in the 2D activation-evaluation space.
In the case of mixing universal emotions, Method 2 combines two universal emotions,
resulting in an intermediate emotion that has features of the two original ones.
114
7.4. VISUALIZATION OF EMOTIONS
Table 7.3: Whissell activation and valuation values, angles and combinations of emotions
as Pi = {Xi,j }, where i is the emotion, j = {1, ..., 64} is the FAP number, and Xi,j
[minV alue, maxV alue]. Also, minV alue [1000, 1000] and maxV alue [1000, 1000].
The objective is to obtain the range of variation Xn,j of each FAP j in En . For this, a
proportionality constant A is computed using Equation 7.2, which is the proportion in
which the activation value of the universal emotion Ei will be present in the intermediate
emotion En . Then, A is multiplied by the range of variation Xi,j of the universal emotion
Ei to get Xn,j , according to Equation 7.3. Xn,j is the translated range of Xi,j due to the
activation value of En . Figure 7.10 shows an example of categorization.
an
A= (7.2)
ai
115
CHAPTER 7. VISUALIZATION OF AFFECT IN FACES
The second method for generating En is applied when it is between two universal emotions
that can be considered close to each other. Then we have to combine the FAPs of both
universal emotions according to the following rules:
Rule I:
If FAP j is involved in the set of activated FAPs P1 , P2 with the same sign (which
represents the same direction of movement):
(a) First we compute the weighted translations of X1,j and X2,j , according to Equa-
tions 7.2 and 7.3. Let these translations be t(X1,j ) and t(X2,j ), which represent
sub-ranges of the FAPs of E1 and E2 . Figure 7.11 is a representation of how the
range is translated.
116
7.4. VISUALIZATION OF EMOTIONS
(b) Then, we compute the center c1,j , c2,j and the length s1,j , s2,j of the translated
ranges t(X1,j ) and t(X2,j ). The center is the mean between minV alue y maxV alue
of the translated ranges. The length is the number of elements per range.
(c) In order to compute the center cn,j and length sn,j of En , first we need to compute
the angles 1 and 2 for E1 and E2 using Equation 7.1.
(d) Using the angles obtained in (c), we compute two constants 1 and 2 . 1
indicates the proportion of E1 E2 in En E1 . 2 indicates the proportion of E1 E2
in En E2 . This is done using Equations 7.4 and 7.5.
n 1
1 = (7.4)
2 1
2 n
2 = (7.5)
2 1
Figure 7.12 shows that the values represented by 1 and 2 are the proportion E1
needs to move to get to En .
(e) Finally, using Equations 7.6 and 7.7 we compute the center and length of the
ranges of variation of En multiplying 1 and 2 by the center and length of the
translated ranges or E1 and E2 . With these values we can compute the final variation
range for En .
117
CHAPTER 7. VISUALIZATION OF AFFECT IN FACES
Rule II:
If FAP j is involved in both P1 , P2 but with contradictory sign, then Xn,j is null
because FAPs are cancelled given that an intersection between the ranges of E1 and
E2 is never found.
Rule III:
If FAP j is involved only in one of P1 or P2 , then the range of variation Xn,j will be
averaged with the ranges of the neutral face, according to Equation 7.2 and 7.8.
Xi,j
Xn,j = A (7.8)
2
As a result of these three rules, we obtain a set of FAPs with a range of variation
that indicates all the possible intensities of the emotion. Figure 7.13 shows the expression
obtained by combining the universal emotions joy and surprise. Figure 7.14 shows some
facial expressions obtained for intermediate emotions.
The methods previously explained can also be applied to other standards as demonstrated
by Albrecht et al. [1], who modified the work of Raouzaiou et al., aimed at an MPEG-4
based model, to a physics-based facial animation system.
In the same fashion, and to prove that the algorithm can be implemented independently
of the animation system, we use FACS to generate intermediate emotions from universal
emotions and the Game Engine of the University of Augsburg.
118
7.4. VISUALIZATION OF EMOTIONS
For the universal emotions, we followed the same methodology as with FAPs. In this
case, we use the Facial Expression Repertoire (FER) [20] to obtain the AUs corresponding
to each expression, as shown in Figures 7.15 and 7.16.
To generate intermediate emotions using FACS we also use the two methods applied
with MPEG-4.
Let E1 , E2 y En be the two universal emotions 1 and 2, and the intermediate emotion
n; a1 , a2 , an , e1 , e2 , en are the centered activation and evaluation values for each emotion.
The set of activated AUs in each emotion is given by: P1 , P2 , Pn . Each Pi has a range of
variation Xi,j , such as Pi = {Xi,j }, where i is the emotion, j is the number that identifies
the AU according to the FACS manual [48], and Xi,j [0, 1]. The main difference with
119
CHAPTER 7. VISUALIZATION OF AFFECT IN FACES
the methodology used with MPEG-4 is that in this case Xi,j is considered a single value,
the intensity of the AU j.
Figure 7.17show a set of facial expressions obtained for intermediate emotions.
Figure 7.17: Upper row: Love, Disappointment. Lower row: Hate, Pity
120
7.5. VISUALIZATION OF MOOD
The methodology to map AUs into PAD is: (1) mapping of emotions into the PAD
space, (2) AUs analysis of facial expressions of emotions using the Facial Expression Reper-
toire (FER) and (3) AUs mapping in PAD Space.
The ALMA model [63] and the work of Russell and Mehrabian [136] provides us with the
PAD values for a set of emotions. These values are in Table 7.4 and some are represented
in Figure 7.19
121
CHAPTER 7. VISUALIZATION OF AFFECT IN FACES
Emotion P A D Mood
Admiration 0.4 -0.49 -0.24 +P-A-D Docile
Anger -0.51 0.59 0.25 -P+A+D Hostile
Arrogance 0.0 0.34 0.5 -P+A+D Hostile
Confusion -0.53 0.27 -0.32 -P+A-D Anxious
Disliking -0.4 -0.2 0.1 -P-A+D Disdainful
Disappointment -0.3 -0.4 -0.4 -P-A-D Bored
Rage -0.44 0.72 0.32 -P+A+D Hostile
Sadness -0.5 -0.42 -0.23 -P-A-D Bored
Fear -0.64 0.60 -0.43 -P+A-D Anxious
Gloating 0.3 -0.3 -0.1 +P-A-D Docile
Gratification 0.6 -0.3 0.4 +P+A+D Exuberant
Gratitude 0.2 0.5 -0.3 +P+A-D Dependent
Hate -0.4 -0.2 0.4 -P-A+D Disdainful
Hope 0.2 0.2 -0.1 +P+A-D Dependent
Joy 0.5 0.42 0.23 +P+A+D Exuberant
Liking 0.40 0.16 -0.24 +P+A-D Dependent
Love 0.3 0.1 0.2 +P+A+D Exuberant
Pity -0.4 -0.2 -0.5 -P-A-D Bored
Pride 0.52 0.22 0.61 +P+A+D Exuberant
Relief 0.2 -0.3 0.4 +P-A+D Relaxed
Resentment -0.3 0.1 -0.6 -P+A-D Anxious
Reproach -0.3 -0.1 0.4 -P-A+D Disdainful
Remorse -0.2 -0.3 -0.2 -P-A-D Bored
Satisfaction 0.50 0.42 0.47 +P-A+D Exuberant
Shame -0.3 0.1 -0.6 -P+A-D Anxious
Terror -0.62 0.82 -0.43 -P+A-D Anxious
Worry (insecure) -0.57 0.14 -0.42 -P+A-D Anxious
Fatigue -0.18 -0.57 -0.29 -P-A-D Bored
The AUs that describe the different emotional expressions are obtained from the Facial
Expression Repertoire (FER) [20]. It is an on-line database developed by the Filmakademie
Baden-Wurttemberg, which maps over 150 emotional expressions to FACS and explains in
detail which AUs must be activated for certain facial expressions. It provides an extensive
data set with different expressions and their corresponding AUs, that will be used to create
the required expressions.
The expressions we use for the analysis are presented in Table 7.5.
122
7.5. VISUALIZATION OF MOOD
(a) Enraged (b) Embarrassment smile (c) Ingratiating Smile (d) Contempt
(e) Arrogance (f) Qualifier smile (g) Coy smile (h) Fear
(q) Tired (r) Disdain (s) Relief Happiness (t) Slight Sadness
123
CHAPTER 7. VISUALIZATION OF AFFECT IN FACES
Given that an emotion is located in certain PAD region, and that this emotion has a set of
AUs, we need to figure out the correspondence between a PAD region and all of the AUs.
The methodology consists in using the emotions in Table 7.4 and identify the AUs
that describe their movements. This gives us a rough idea of where the AUs are mainly
active. The difficulty lies in deciding the PAD region where the AU should be active when
it belongs to emotions in different octants. For example, AU 1 is active in fear which is in
the -P -D quadrant, but is also active in other emotions in the +P -D quadrant.
The analysis is divided in: analysis of the AUs in the Pleasure-Dominance (PD) space,
and analysis of the AUs in the Arousal (A) space. One of the reasons for this divisions is
that it facilitates the analysis of the dimensions where the AUs could be activated. Another
reason is that there are AUs that can be easily associated with arousal or activation, as
opening of the mouth or the eyes.
From these analysis we can formulate a function dependent on the Pleasure-Dominance
or Arousal dimension for each AU, which depicts the area where the AU is active.
As for the analyzed AUs, we consider a reduced set of AUs that results potentially
sufficient to express, in a readily recognizable manner a set of facial expressions [53]. The
final set of AUs used in this research are presented in Table 7.6.
124
7.5. VISUALIZATION OF MOOD
In the following we present the emotions and methodology used in the analysis of each
AU. Table 7.8 contains all the equations to map each AU into the different dimensions.
125
CHAPTER 7. VISUALIZATION OF AFFECT IN FACES
To map the AU 1 into the +P-D quadrant using relief (p = 0.2 and d = 0.4)
according to Equation 7.10, which is similar to Equation 7.9, but with an offset of
0.4 because relief is the point where AU 1 begins to change.
0.0 if d [0.4, 0.0)
AU 1intensity = (4.0)(d + 0.4) if d (0.65, 0.4) (7.10)
1.0 if d [1.0, 0.65]
For the -P-D quadrant, we use sadness and fear to compute the area. It results in
a progression beginning at the location of sadness (p = 0.5 and d = 0.25), and
ending where fear is located (p = 0.65 and d = 0.45).
For the +P-D quadrant we used the same analysis as for AU 1. Figure 7.21 shows
the area where AU 2 is activated. The functions to obtain the areas are in Table 7.8.
126
7.5. VISUALIZATION OF MOOD
AU 4 - Brow Lowerer
AU 4 is found in the negative Pleasure (-P) along the Dominance dimension.
To map this AU in the -P+D quadrant we use anger (p = 0.5 and d = 0.25)
because it is located in that quadrant. In FER the expressions considered as anger
variants are: enraged (compressed lips) and sternness, which have about the same
dominance value on the -P side (Figure 7.5).
For the -P-D quadrant we use sadness (p = 0.5 and d = 0.25) to compute
the intensity of AU 4. Figure 7.22 shows the area where AU 4 is activated and the
equations to obtain this area are shown in Table 7.8.
127
CHAPTER 7. VISUALIZATION OF AFFECT IN FACES
AU 6 - Cheek Raiser
The feature raised cheeks is a key element for joy and appears in displays of genuine
emotion, but is missing in fake smiles masking other feelings. Therefore, we map this
AU to the positive Pleasure (+P) along the Dominance dimension.
AU 6 is seen in the FER expressions for embarrassment smile and ingratiating smile
(Figure 7.5), which are representative for shame. Nevertheless, shame tends not to
be represented with a smile. Therefore, we concluded that the mentioned expressions
of smile are a sign of agreeableness towards the interlocutor. That is why we use
liking (p = 0.4 and d = 0.24) for the negative dominance (+P-D).
For the positive dominance (+P+D) we use the emotions joy (p = 0.5 and d = 0.25)
and satisfaction (p = 0.5 and d = 0.47). Figure 7.23 shows the area where it is
activated and Table 7.8 shows the equations to obtain the area of activation.
128
7.5. VISUALIZATION OF MOOD
AU 14 - Dimpler
This AU appears in a smiling mouth. This marks a key feature for the positive
Pleasure dimension (+P).
One of the FER expressions described by this AU is enjoyable contempt (Figure 7.5),
which can be seen as the positive pleasure variety of arrogance.
Thus, AU 14 is mapped to the +P+D quadrant. However, as its region of activation
is closer to the maximum dominance values and lower positive pleasure values, for
computation of this AU we take into account both P and D. Figure 7.26 shows the
area where AU 14 is activated.
129
CHAPTER 7. VISUALIZATION OF AFFECT IN FACES
130
7.5. VISUALIZATION OF MOOD
131
CHAPTER 7. VISUALIZATION OF AFFECT IN FACES
AU 25 - Lips Part
According to FER, the AU 25 is found in emotions like rage, surprise, disgust, or
in expressions like dazzled smile, which would correspond to a confused expression
(Figure 7.5). All these emotions have distinctive values in the positive Arousal
dimension.
We use confusion (a = 0.27) and rage with a = 0.72 to compute the AUs activation.
Figure 7.29 shows how the AU increased its intensity only in the activation dimension.
AU 26 - Jaw Drop
The AU 26 describes a similar movement as AU 25, but it involves a movement of the
bone, which might result sometimes in a more exaggerated opening of the mouth.
Nevertheless, AU 26 is also found in surprise, fear or disgust, among other FER
expressions like uproarious laughter and yawning (Figure 7.5).
We use the values of fear (a = 0.6) and disgust, (a = 0.35) to compute this AU
intensity. Figure 7.30 shows how the AU increased its intensity only in the activation
dimension.
132
7.5. VISUALIZATION OF MOOD
AU 43 - Eye Closure
AU 43 describes a movement that is associated emotions or states like tiredness, dis-
dain or relief happiness (Figure 7.5). Thus it belongs to the negative Arousal
dimension (-A).
To compute the area where AU 43 is activated we use the arousal value of fatigued
(a = 0.57). Figure 7.31 shows the location of this AU in the activation dimension.
As a result, the set of AUs in each mood are shown in Table 7.7.
Table 7.8 is a summary of all the equations formulated with the previous analysis, and
it also contains the graphs that depict the areas where the AUs are activated in the PAD
space.
133
134
AU Emotions Equations Comments
(4.0)d; d (0.25, 0.0] p [1.0, 0.0)
sadness AU 1 = Linear function from d = 0.0
1.0; d [1.0, 0.25] p [1.0, 0.0)
AU1 to d = 0.25
0.0; d [0.4, 0.0) p (0.0, 1.0]
relief AU 1 = (4.0)(d + 0.4); d (0.65, 0.4) p (0.0, 1.0] offset of -0.4 because AU 1
1.0; begins to change in relief
d [1.0, 0.65] p (0.0, 1.0]
0.0; d (0.25, 0.0] p [1.0, 0.0)
sadness, fear AU 2 = (4.0)d; d (0.45, 0.25] p [1.0, 0.0) progression begins at sad-
AU2
1.0; ness, ending at fear
d [1.0, 0.45) p [1.0, 0.0)
0.0; d [0.4, 0.0] p (0.0, 1.0]
relief AU 2 = (4.0)(d + 0.4); d (0.65, 0.4) p (0.0, 1.0]
1.0; d [1.0, 0.65] p (0.0, 1.0]
(4d)(2p); p (0.5, 0.0] d (0.0, 0.25)
(2.0)p; p (0.5, 0.0] d [0.25, 1.0]
anger AU 4 = this AU is dependent on
(4.0)d; p [1.0, 0.5] d (0.0, 0.25)
AU4
pleasure and dominance
1.0; p [1.0, 0.5] d [0.25, 1.0]
(4d)(2p); p (0.5, 0.0] d (0.25, 0.0]
(2.0)p; p (0.5, 0.0] d (1.0, 0.25]
AU 4 =
(4.0)d; p [1.0, 0.5] d (0.25, 0.0]
1.0; p [1.0, 0.5] d [1.0, 0.25]
(2.0)(d + 0.25); d (0.25, 0.25)
liking, joy, pride AU 6 = (4.0)(0.5 d); d [0.25, 0.5)
AU6
1.0; d [0.5, 1.0]
(2.0)(d 0.5); d [0.5, 1.0]
arrogance AU 10 =
AU10 0.0; d [0.0, 0.5)
CHAPTER 7. VISUALIZATION OF AFFECT IN FACES
0.0; p [1.0, 0.0)
joy AU 12 = (2.0)p; p [0.0, 0.5) Joy is representative along
AU12
1.0; dominance
p [0.5, 1.0]
2(0.5 p)2(d 0.5); p [0.0, 0.5) d (0.5, 1.0]
enjoyable contempt AU 14 = Enjoyable contempt: arro-
AU14 0.0; else gance with positive pleasure
0.0; p (0.0, 1.0]
sadness AU 15 = (2.0)p; p (0.5, 0.0]
AU15
1.0; p [1.0, 0.5]
0.0; a [1.0, 0.1]
7.5. VISUALIZATION OF MOOD
Table 7.8: Equations to obtain the degree of activation of each AU in the PAD space
135
CHAPTER 7. VISUALIZATION OF AFFECT IN FACES
136
7.6. VISUALIZATION OF PERSONALITY TRAITS
their influence when a virtual agent is expected to express social dominance; and Arya et
al. [7] who defined a set of visual cues (among them head turn, head tilt, head down and
averted gaze) to map personality dimensions Dominance and Affiliation.
Finally, we decided to also study neuroticism for being the other personality trait that
is present in all personality models, from Eysenck [52] to these days.
7.6.2 Hypothesis
In the experiment that is going to be explained in the following, we assess the relationship
between these non-verbal facial cues and their influence on the perception of a characters
personality. We expect the following outcomes:
The perception of the three personality traits Extraversion, Agreeableness, and Emo-
tional Stability is not influenced whether the virtual characters head points to the
left or to the right.
Depending on the personality trait, direction plays a role in how these traits are
perceived. We expect that it makes a difference, for example, how Extraversion is
perceived in contrast to Agreeableness when the character is looking upwards.
Not only head orientation influences the perception of the personality traits. Also
variations of eye gaze directions further influence how the personality traits of the
virtual character are perceived.
7.6.3 Methodology
1. Creating different static images with combined head poses (upwards, center, down-
wards, sideways) and different eye gaze (upwards, center, downwards, sideways)
137
CHAPTER 7. VISUALIZATION OF AFFECT IN FACES
To carry out the experiment we used the virtual agent Alfred, from the Game Engine of
the University of Augsburg 7.3.1.
The visual cues: head movement and gaze orientation were calculated by varying hor-
izontal and vertical angles, each in three symmetric steps.
For both vertical and horizontal axis, the neutral center position remained at 0.0. The
positions for turning the head sideways were set as 8.5 for looking left, from the agents
point of view, and -8.5 for looking right. For tilting the head vertically, the high target
was located at 8.0 degrees, whereas the low one was at -8.0. Since the distance between
neck and eye joints weakened the visible effect on the eye movement, the vertical angles
had to be doubled for the eye targets.
All limits were chosen in regard to the goal that the pupils should remain visible, even
when the eyes look in the opposite direction of the head.
By combining these angles, nine different targets could be provided for the survey.
These were then converted to Cartesian coordinates using a fixed radius for all target
angles, and sent to the virtual agents IK component for every combination of head pose
and eye gaze. The 81 resulting expressions were captured as screenshots.
However, to ensure a sufficient number of votes per picture, the number of samples had
to be reduced and redundant combinations eliminated. To do this, previous observations
were performed with a reduced group of users, obtaining as a result that the direction of
lateral head movements would not cause much of a difference. Thus we decided to merge
both left and right looking images into one sideways category. To keep the natural
variation, about half of the required images were chosen randomly to either gaze in one
or the other direction. The associated eye gaze targets were mirrored to keep the proper
relation between head and eye movements.
In the end, we worked with a reduced set of 54 images (6 head directions 9 eye
directions) of Alfred. Combinations of orientations were written: vertical -horizontal, e.g.
upwards-center. Table 7.9 shows some samples of head orientations.
138
7.6. VISUALIZATION OF PERSONALITY TRAITS
head down
Questionnaire
133 subjects (47 female and 86 male) participated in the experiment through an online
questionnaire. The mean age was 26.6 (SD = 8.8). The questions were provided in
English, German or Spanish, depending on the subjects mother tongue.
The questionnaire consisted of 54 static images, where each image was judged at least 10
times. The images corresponded to a virtual character in which head orientation (upwards-
center, upwards-sideways, middle-center, middle-sideways, downwards-center, downwards-
sideways) and eye gaze (upwards-center, downwards-center, upwards-sideways, downwards-
sideways) were combined.
Then the experimental stimuli which consisted of 15 images per user were presented
one at a time, in random order. For each stimulus the participant had to answer to six
items of the Ten-Item Personality Inventory (TIPI) [66] presented in a 7-item Likert Scale,
where 1 corresponded to Disagree Strongly and 7 to Agree Strongly. Table 7.10 shows
the items presented for each image of the questionnaire.
139
CHAPTER 7. VISUALIZATION OF AFFECT IN FACES
Trait Item
Extraversion Extraverted, enthusiastic
Reserved, quiet
Agreeableness Critical, quarrelsome
Sympathetic, warm
Emotional Stability Anxious, easily upset
Calm, emotionally stable
7.6.5 Results
The following results show the mean values and standard deviations related to how
was Alfred perceived over all ratings:
To study if the side where the agent is looking to has an influence on the perception
of personality, we assumed that in general there are no noticeable difference among the
personality traits whether the agent looks to the right or to the left.
The method was to apply a two-tailed independent t-test to the overall values for
Extraversion, Agreeableness, and Emotional Stability dependent on which side the virtual
agent is looking.
The results of the t-test with 329 degrees of freedom for the trait of Extraversion,
t(329) = 1.2, p = .25, r = .07, showed that there was no significant difference for
Extraversion between Alfred looking to the left ( = 3.8, = 1.2) and looking to the right
( = 3.9, = 1.3), given that the obtained p-value is above the probability threshold.
Moreover, the effect size r has a value below .1, which demonstrates the weak relationship
between Extraversion and the side to where the character is looking to. Therefore, we can
accept the assumption of no noticeable difference in Extraversion when the agent looks to
the left or to the right.
140
7.6. VISUALIZATION OF PERSONALITY TRAITS
Agreeableness also did not show any significant differences for the virtual character
Alfred between looking to the left ( = 3.8, = 1.5) and looking to the right ( = 3.7,
= 1.4), t(329) = .62, p = .54, r = .03. Again, given that the obtained p-value is above
the threshold and the effect size r is below .1, a weak relationship between Agreeableness
and the side to where the character is looking to is shown. Hence, we also can accept the
assumption of no noticeable difference.
Finally, Emotional Stability as well did not show any significant differences between
looking to the left ( = 4.4, = 1.4) and looking to the right ( = 4.4, = 1.2),
t(329) = .35, p = .73, r = .02. In the latter case, as with Extraversion and Agreeableness,
the effect size r is for all three personality dimensions below .1 and thus can be further
interpreted as not even a small effect.
Extraversion
Positioning the head to the upwards-sideways got the highest rating for Extraversion,
while redirecting the head downwards-center got the lowest rating for Extraversion
(Figures 7.32 and 7.33).
Figure 7.32: Mean values for Extraversion dependent on the head orientation.
The one-way ANOVA showed that there was a significant effect on the perception
of Extraversion on levels of the different head orientations, F (5, 663) = 15.4, p < .001,
2 = .10.
Tukey post hoc tests revealed several significant differences within the perception of
Extraversion dependent on the head orientation. Table7.11 shows these results, where each
row and column corresponds to the combination of vertical and horizontal positioning, e.g.
U-C means upwards-center.
141
CHAPTER 7. VISUALIZATION OF AFFECT IN FACES
Figure 7.33: The head orientation (downwards-center ) with the lowest rating (left) and the
one (upwards-sideways) with the highest rating (right) for Extraversion.
Table 7.11: Post-hoc comparisons for Extraversion and the varying head orientations.
Vertical orientations: U = upwards, M = middle, D = downwards. Horizontal orientations:
S = sideways, C = center. + p < .1, * p < .05, *** p < .001, n.s. = not significant.
Alfred with its head pointing upwards-sideways ( = 4.3, = 1.2) was perceived
significantly less extraverted than the heads pointing to the middle-sideways ( = 3.8,
= 1.3, p < .05), to the middle-center ( = 3.5, = 1.2, p < .001), downwards-side
( = 3.4, = 1.1, p < .001), and downwards-center ( = 3.0, = 1.1, p < .001).
An upwards-center head position ( = 3.9, = 1.2) is perceived as less extraverted
than the heads looking to the center ( = 3.5, = 1.2, p < .1), downwards-side ( = 3.4,
= 1.1, p < .05), and downwards-center ( = 3.0, = 1.1, p < .001).
A head directed middle-sideways ( = 3.8, = 1.3) is perceived as less extraverted
than a head looking downwards-center ( = 3.5, = 1.2, p < .001).
As we applied a two-tailed post hoc test, the significant results are also valid vice versa.
In general we can see that when Alfred has the head sideways, the perception of Ex-
traversion is increased independently of the vertical orientation (Figure 7.32). Nevertheless,
the vertical orientation has also influence in the perception. A raised head is perceived as
142
7.6. VISUALIZATION OF PERSONALITY TRAITS
Agreeableness
The highest value for Agreeableness was achieved when the virtual character looked
downwards-center . The lowest value was achieved for a virtual character looking
upwards-center (Figures 7.34 and 7.35).
Figure 7.34: Mean values for Agreeableness dependent on the head orientation.
Figure 7.35: The head orientation (upwards-center ) with the lowest rating (left) and the
one (downwards-center ) with the highest rating (right) for Agreeableness.
There was a significant effect on the perception of Agreeableness on levels of the different
head orientations, F (5, 663) = 14.4, p < .001, 2 = .09.
Table 7.12 shows the results of the Tukey post hoc tests, which revealed several signifi-
cant differences within the perception of Agreeableness dependent on the head orientation
143
CHAPTER 7. VISUALIZATION OF AFFECT IN FACES
Table 7.12: Post-hoc comparisons for Agreeableness and varying head orientation. Vertical
orientations: U = upwards, M = middle, D = downwards. Horizontal orientations: S =
sideways, C = center. ** p < .01, *** p < .001, n.s. = not significant
The head directed upwards-sideways ( = 3.3, = 1.3) was perceived as less agreeable
than a head directed to the middle-sideways ( = 4.0, = 1.4, p < .001), a centered head
( = 3.9, = 1.3, p < .01), a head downwards-sideways ( = 4.1, = 1.4, p < .001), or
downwards-center ( = 4.3, = 1.2, p < .001).
Alfred with the head looking upwards-center ( = 3.1, = 1.3) was perceived as less
agreeable than when looking to the center-sideways ( = 4.0, = 1.4, p < .001), to the
center ( = 3.9, = 1.3, p < .001), downwards-sideways ( = 4.1, = 1.4, p < .001), and
downwards-center ( = 4.3, = 1.2, p < .001).
The lowest values for Agreeableness were achieved for the virtual character looking
upwards. Higher values could be achieved for Alfred with centered head, and even slightly
higher values were achieved for looking downwards (Figure 7.34).
Regarding eye gaze, also for Agreeableness we could not find any significant differences
for the varying eye gaze directions dependent on the six head orientations.
Emotional Stability
Alfred looking middle-side achieved the highest ratings for Emotional Stability and
the upwards-center orientation achieved the lowest ratings (Figures 7.36 and 7.37).
There was a significant effect on the perception of Emotional Stability on levels of the
different head orientations, F (5, 663) = 3.6, p < .01, 2 = .02.
Tukey post hoc tests revealed only one significant difference within the perception of
Emotional Stability dependent on the head orientation.
Alfreds head directed to the middle-sideways ( = 4.7, = 1.2) was perceived with
significantly lower Emotional Stability than directed upwards-center ( = 4.0, = 1.3)
144
7.6. VISUALIZATION OF PERSONALITY TRAITS
Figure 7.36: Mean values for Emotional Stability dependent on the head orientation.
Figure 7.37: The head orientation (upwards-center ) with the lowest rating (left) and the
one (middle-sideways) with the highest rating (right) for Emotional Stability.
7.6.6 Discussion
The obtained results provided us with data that could be used to improve the modeling
of personality in virtual agents and therefore, the interaction between real users and these
agents. An important aspect was the study of visual cues for certain personality traits that
145
CHAPTER 7. VISUALIZATION OF AFFECT IN FACES
7.7 Summary
This chapter described the algorithms and methods we used to visualize emotions, mood
and personality.
Emotions were visualized in a MPEG-4 based avatar, in which we implemented an
algorithm that using expressions of universal emotions is capable of generating expressions
for intermediate emotions, either by combining two universal emotions or categorizing one
of them. As a result we obtained a set of facial expressions which were correctly identified
by a group of subjects, demonstrating the validity of the algorithm.
One of the novelties of this work is the generation of facial expressions for moods. Our
main contribution is the development of a set of FACS-based functions which locate AUs
into the PAD Space model, allowing us to associate those AUs with the moods correspond-
ing to the PAD model. In this way, we can generate facial expressions for the moods using
the AUs activated in the corresponding octants of the space.
Finally, for personality we performed an experiment to explore how the visual cues:
head pose and eye gaze, influence the perception of three personality traits: extraver-
sion, agreeableness and emotional stability. In result we obtained that for the evaluated
character, extraversion was associated with a head in an upwards-sideways position, agree-
ableness with the head in a downwards-center position, and emotional stability with a
center-sideways position. Using this subjective information about how personality is per-
ceived we could enhance facial expressions, achieving more believable characters.
146
Chapter 8
Evaluation
Not everything that can be counted counts, and not everything that counts can be counted.
Albert Einstein
The results of formal evaluations are presented in this chapter aiming at assessing the
degree of users perception, the believability of the characters and the effectiveness of the
computational affective model we have presented.
To achieve our objective, we performed a number of individual experiments with dif-
ferent subjects. That was the most accurate way to evaluate each part of the framework
and guarantee its functionality, given that a quantitative evaluation of the whole system
was infeasible.
The chapter is organized as follows. First, we present the general objectives of this
evaluation and what we want to achieve. Then, the following subsections explain in detail
the subjects, apparatus and procedures used in each experiment. Finally, a summary of
the chapter is given.
147
CHAPTER 8. EVALUATION
which takes the emotions elicited by the semantic model, maps them into a Pleasure-
Arousal-Dominance space (PAD) and process them to obtain the mood of the character.
Finally, the third module is the visualization module, which implements the algorithms
and methods for the generation of facial expressions of emotions and mood.
Nevertheless, an evaluation of the whole framework, all in once is a difficult task because
there would be too many variables to take into consideration. That is why we broke down
the evaluation process and designed various experiments which will validate our work and
will give us hints of how to improve the generation of context and the visualization of
affective traits. The evaluation objectives are:
1. Validation of the expressions of basic and intermediate emotions, and their perception
as such by the subjects.
2. Validation of the expressions of mood elicited by the Affective Module, and their
perception as such by the subjects.
3. Evaluation of how the emotions elicited in different scenes of the Context Represen-
tation Module are perceived by the subjects.
In the following we explain how we achieve the former objectives through subjective
experiments.
8.2.1 Hypothesis
Before carrying on this experiment we formulated the following hypothesis:
Given that intermediate emotions are obtained from the combination of two universal
ones, subjects would be capable of identify at least one of the universal emotions that
composed the intermediate one.
148
8.2. EXPERIMENT: VISUALIZATION OF EMOTIONS
Regarding the second hypothesis, we did not ask for recognition of the specific inter-
mediate emotion because, as Ekman mentioned in [48], the name one can give to an
emotion depends a lot on context and the situation in which that emotion arises.
8.2.2 Methodology
To perform the evaluation, first we selected a group of images correspondent to a set of the
generated facial expressions. This selection was the result of previous observations, where
we realized that different emotions shared the same facial expression. Table 8.1 contains a
list with the emotions which expressions were evaluated.
To evaluate the validity of the generated expressions, we carried out a subjective study
with the following methodology:
149
CHAPTER 8. EVALUATION
Questionnaire
Results
Table 8.2 shows the results of recognition of universal emotions on each of the 16 expres-
sions. First column presents the emotion associated to the evaluated expression. Second
column presents the emotions that are combined to generate the corresponding interme-
diate emotion. The following columns present the percentages of the number of hits (%),
regarding recognition of the facial expressions.
150
8.2. EXPERIMENT: VISUALIZATION OF EMOTIONS
Universal Emotions. 6 expressions for universal emotions were studied. Figure 8.1
depicts the recognition hits percentages presented in Table 8.2.
151
CHAPTER 8. EVALUATION
As can be seen in Figure 8.1, in most of the cases expressions of joy, sadness, anger
and surprise where recognized. On the other hand, disgust and fear obtained lower
success rate. Disgust tended to be confused with anger, and fear with surprise.
This last result has to do with the fact that our expressions for surprise and fear
share some facial configurations as raised eyebrows and open mouth. Nevertheless,
this result was expected because both expressions are mainly differentiated through
context where the event is taking place and timing issues, which cannot be represented
on a still image.
Moreover, Ekman and Friesen [47] see fear differing from surprise in three ways:
1. Whilst surprise is not necessarily pleasant or unpleasant, even mild fear is un-
pleasant.
2. Something familiar can induce fear, but hardly surprise (for example, a visit to
the dentist).
3. Whilst surprise usually disappears as soon as it is clear what the surprising event
was, fear can last much longer, even if the event is known.
8.2.4 Discussion
At the beginning of this section two hypothesis were formulated. In the first hypothesis we
wanted to see if expressions of universal emotions were better perceived than expressions of
152
8.2. EXPERIMENT: VISUALIZATION OF EMOTIONS
Figure 8.2: % of recognition success for Intermediate emotions categorized from 1 universal
emotion
Figure 8.3: % success for Intermediate emotions generated from combining 2 universal
emotions
153
CHAPTER 8. EVALUATION
intermediate emotions. From the results obtained we could observe that universal emotions
were indeed easily recognized. This fact helped to the good results in the recognition of
intermediate expressions, given that the universal ones we used to generate them were
correctly assessed.
Regarding the second hypothesis, we wanted to know if subjects could identify the uni-
versal emotion (or emotions) used for the generation of the expression of the intermediate
emotion. Indeed, it was proved with the results that showed that the great majority of
the participants recognized in an high percentage the universal emotion from which the
intermediate is obtained. Even more, the results showed that when it came to intermediate
expressions generated from two universal ones, the two corresponding universal emotions
obtained the highest recognition rates above the rest of universal emotions.
154
8.3. EXPERIMENT: VISUALIZATION OF MOODS
8.3.1 Hypothesis
The generation of facial expressions for moods is a topic that has been poorly addressed
by previous researches. Therefore, this evaluation is a mixing of experimentation and
validation, expecting to prove the following hypothesis:
All 8 moods have corresponding facial expressions, which are described by the AUs
activated in the PAD octant of that mood according to the rules formulated in Sec-
tion 7.5.
8.3.2 Methodology
The methodology of this experiment goes from the creation of the expressions to evaluate
to the analysis of the results. The methodology steps are:
1. Generation of a set of images per each mood. Given that each mood has 3 dimensions
(pleasure, arousal and dominance) and 3 intensity values per dimension (low, medium,
high), in total (8)(33 ) = 216 images were generated.
3. Evaluation of the results to verify that expressions associated to each mood exist,
and to obtain the AUs that describe those expressions.
155
CHAPTER 8. EVALUATION
Table 8.3: Facial expressions in the mood quadrants of the PAD Space. Upper row:
Anxious, Bored, Dependent, Disdainful. Lower row: Docile, Exuberant, Hostile, Relaxed
We evaluated the expressions based on their PAD values and not on their adjectives,
because we wanted to do it in a simpler manner that could be understood across cultures,
avoiding the issue of translating each adjective and still maintain its correct meaning.
Questionnaire
109 subjects (59 male and 50 female) between 19 and 55 years old, with a mean age of 29.2
(SD = 7.1) participated in the experiment through an online questionnaire.
To assess the images we used the Self-Assessment Manikin questionnaire, which expla-
nation was provided in English and Spanish, so we could get a greater sample of subjects.
The Self-Assessment Manikin (SAM) [22] is a non-verbal, graphic representation of the
156
8.3. EXPERIMENT: VISUALIZATION OF MOODS
three dimensions: Pleasure, Arousal and Dominance. It directly assesses the pleasure,
arousal and dominance associated in response to an object or event. SAM ranges from
a smiling, happy figure to a frowning, unhappy figure when representing the pleasure
dimension; from an excited, wide-eyed figure to a relaxed, sleepy figure for the arousal
dimension; and from a small to a large figure that indicates maximum control in the
situation in the case of the dominance dimension.
The experimental stimuli consisted of 18 static images, randomly selected from the pool
of 216 images. Then, each subject had to rate each of these 18 images using SAM, which
was presented in a 5-item Likert Scale, where 1 corresponded to the minimum value of
the dimension and 5 to the maximum. For analysis purposes, this scale was normalized
between 1 and 1. The questions were of the form:
(1) How is Alfred feeling?. Possible answers corresponded to the SAM items for pleasure:
very displeased (-1), displeased (0.5), neutral (0.0), pleased (0.5), very pleased (1.0).
(2) How energetic seems Alfred?. Possible answers corresponded to the SAM items for
arousal: very relaxed (-1), relaxed (0.5), neutral (0.0), excited (0.5), very excited (1.0).
(3) How dominant is Alfred?. The possible answers corresponded to the SAM items
for arousal: very submissive (-1), submissive (0.5), neutral (0.0), dominant (0.5), very
dominant (1.0).
Figure 8.4 shows a page of the online questionnaire.
157
CHAPTER 8. EVALUATION
The first two images presented to the subjects were for training purposes. They corre-
sponded to the moods Exuberant and Bored with their maximum values: +P+A+D and
-P-A-D, respectively. Thus the subject could evaluate these images using the SAM figures,
and then continue with the remaining 16 random images.
8.3.4 Results
The obtained results opened the possibility to associate facial expressions to moods defined
by the values of pleasure, arousal and dominance.
When analyzing the results, we could observe that each image was evaluated among
5 and 13 times (with a mean value 10). The explanation for this variability is the
randomness of the image generation.
Table 8.4 presents the mean values and standard deviations after analyzing each PAD
dimension in all images corresponding to one of the eight moods. Figure 8.5 shows graphi-
cally the mean values for each dimension for each mood. Blue bars correspond to pleasure
(P) values, red bars to arousal (A) and green bars to dominance (D).
Given that the three dimensions pleasure, arousal and dominance are independent, we
need to analyze each image to obtain the best combination of dimensions intensities, and
hence to obtain the significant image(s) representing each mood.
In the following, a detailed analysis is presented, so we can know exactly which expres-
sions were recognized as representatives for each mood.
158
8.3. EXPERIMENT: VISUALIZATION OF MOODS
Mood: Anxious
This mood corresponds to the -P+A-D quadrant of the PAD model. Therefore, we expect
a set of expressions corresponding to this mood to be recognized as with negative pleasure
and dominance, and positive arousal.
Figure 8.6 presents the distribution of the mean recognition values for each of the 27
images corresponding to this mood
Figure 8.6: Mean values of the analysis of facial expressions of mood anxious
Table 8.5 presents the mean values and standard deviations of the expressions rec-
ognized as anxious. The mean values representing the average recognition rate of each
159
CHAPTER 8. EVALUATION
dimension, and the standard deviations representing its variability showed that a set of six
images were identified as -P+A-D.
Table 8.5: Mean values and Standard Deviations for expressions of Anxiousness
The former results constrain our analysis of the images, thus we could assess its validity
through the recognition hit rates for each image. Table 8.6 shows the hit rates of the images
that obtained higher scores and the number of subjects that evaluated each image. The
expressions discarded are those in which their P, A or D was wrongly identified by more
than half of the subjects.
From the previous results we could observe that the images identified in the same
quadrant of anxious have:
- Negative Pleasure intensities along the negative dimension
- Positive Arousal intensity in its maximum level
- Negative Dominance intensity between its medium and maximum level
Using these results we can see which AUs were activated according to the in-
tensities of pleasure, arousal and dominance of the selected expressions, resulting in
AU 1, AU 2, AU 4, AU 15 for pleasure and dominance and AU 5, AU 25, AU 26 for arousal.
These AUs correspond to the ones in the quadrant of anxiousness. Figure 8.7 shows the
validated faces for this mood.
160
8.3. EXPERIMENT: VISUALIZATION OF MOODS
Figure 8.7: Facial expressions with their corresponding PAD intensities for anxious
Mood: Bored
This mood corresponds to the -P-A-D quadrant of the PAD model. Therefore, we expect
expressions corresponding to this mood to be recognized as with negative pleasure, arousal
and dominance.
Figure 8.8 presents the distribution of the mean recognition values of each of the 27
images corresponding to the mood bored.
Figure 8.8: Mean values of the analysis of facial expressions of mood bored
Table 8.7 presents the mean values and standard deviations of the expressions recog-
nized as bored. These results show that most of the images were identified as -P-A-D.
The former results constrain our analysis of the images, thus we could assess its validity
through the recognition hit rates for each image. Table 8.8 shows the hit rates of the images
that obtained higher scores and the number of subjects that evaluated each image. The
161
CHAPTER 8. EVALUATION
Table 8.7: Mean values and Standard Deviations for expressions of Boredness
expressions discarded are those in which their P, A or D was wrongly identified by more
than half of the subjects.
From the previous results we observed that the images have these characteristics:
- Negative pleasure intensity is between high and medium
- Negative arousal and negative dominance intensity can have any degree (high, medium
or low)
It is worth noting that the images that were not recognized as bored present low negative
pleasure (almost neutral regarding pleasure), as can be seen in Figure 8.8. Also, the hit
rates analysis resulted in a reduced set of expressions in compare to the ones obtained
through mean values, mainly because the majority of subjects could not associate an
arousal or dominance value to the expression.
Using these results we can see which AUs were activated according to the intensities
of pleasure, arousal and dominance of the selected expressions, resulting in AU 1, AU 2
(activated only when p > 0.5), AU 4, AU 15 for pleasure and dominance and AU 43 for
arousal. These AUs correspond to the ones in the quadrant of bored. Figure 8.9 shows a
subset of the expressions that obtained highest recognition scores.
162
8.3. EXPERIMENT: VISUALIZATION OF MOODS
Figure 8.9: Facial expressions with their corresponding PAD intensities for bored
163
CHAPTER 8. EVALUATION
Mood: Dependent
This mood corresponds to the +P+A-D quadrant of the PAD model. Therefore, we
expect expressions corresponding to this mood to be recognized as with positive pleasure,
positive arousal and negative dominance. Figure 8.10 presents the distribution of the mean
recognition values of the 27 images for this mood.
Figure 8.10: Mean values of the analysis of facial expressions of mood dependent
Table 8.9 presents the mean and standard deviation values of the expressions that were
assessed as +P+A-D.
Table 8.9: Mean values and Standard Deviations for expressions of Dependent
From the previous results, it can be seen that the value of dominance is very close to
zero, or zero. Regarding the remaining images, the dimensions pleasure and arousal were
correctly located in almost 90% of the cases. Nevertheless, dominance was assessed as
negative in only 14.8% of the cases.
This results show the difficulty when assessing dominance in the expressions of the
mood dependent. This issue is also seen when evaluating the recognition hit rates for each
164
8.3. EXPERIMENT: VISUALIZATION OF MOODS
image, where each one obtains significant values for pleasure and arousal, but values close
to zero for dominance, except for the expression 10-De-MHH. This expression is the only
one in which dominance was recognized by half of the subjects. Table 8.10 shows the hit
rates for the subset of the selected images.
Figure 8.11 shows the expressions that could be assessed as being in +P+A-D. The
AUs that describe these expression are AU 1, AU 2 and AU 12, which are activated only for
positive pleasure with medium to high intensity and theoretically for negative dominance
with high intensity; and AU 5, AU 25, AU 26 for positive arousal with any intensity.
Figure 8.11: Facial expressions with their corresponding PAD intensities for dependent
Mood: Disdainful
This mood corresponds to the -P-A+D quadrant of the PAD model. Therefore, we expect
expressions corresponding to this mood to be recognized as with negative pleasure, negative
arousal and positive dominance.
Figure 8.12 presents the distribution of the mean recognition values for expression
corresponding to this mood.
165
CHAPTER 8. EVALUATION
Figure 8.12: Mean values of the analysis of facial expressions of mood disdainful
Table 8.11 presents the mean and standard deviation values of the expressions recog-
nized as with -P-A+D.
Image Pleasure Arousal Dominance
Mean SD Mean SD Mean SD
15-Di-MML -0.56 0.32 -0.38 0.44 0.13 0.58
23-Di-LMM -0.17 0.25 -0.42 0.58 0.25 0.52
Table 8.11: Mean values and Standard Deviations for expressions of Disdainful
From the previous results it can be seen that two expressions were associated with the
dimensions of disdain. Standard deviation values show that there is a great variability
among the subjects rating. Therefore, we constrain the analysis to assess the hit rates for
these two images. Table 8.12 shows the hit rates for the subset of the selected images.
The hit rates analysis demonstrates that from the two expressions identified in the
disdain mood, the second one (23-Di-LMM) presents low rating for pleasure. It is expected
given that the value for pleasure is low, therefore this expression shares some characteristics
of the neutral expression, where the character has the corner lips slightly raised.
166
8.3. EXPERIMENT: VISUALIZATION OF MOODS
Figure 8.13: Facial expressions with their corresponding PAD intensities for disdainful
Mood: Docile
This mood corresponds to the +P-A-D quadrant of the PAD model. Therefore, we
expect expressions corresponding to this mood to be recognized as with positive pleasure
and negative arousal and dominance.
Figure 8.14 presents the distribution of the mean recognition values of each of the 27
images corresponding to the mood docile.
Table 8.13 presents the numeric mean and standard deviation values of those expressions
that were indeed recognized as +P-A-D.
The former results constrain our analysis to these expressions, which were further stud-
ied by obtaining the recognition hit rates for each dimension. An expression is considered
correspondent to the mood docile, and therefore to the +P-A-D quadrant if more than half
of the subjects identified its three dimensions as such. Table 8.14 shows the hit rates for
the subset of the selected images.
167
CHAPTER 8. EVALUATION
Figure 8.14: Mean values of the analysis of facial expressions of mood docile
Table 8.13: Mean values and Standard Deviations for expressions of Docile
The hit rates analysis demonstrates that the pleasure dimension was recognized in
all the cases with a high percentage rate. Also, arousal and dominance were correctly
identified, although in some cases arousal obtained a lower success rate than dominance.
Regarding the expressions that were recognized as docile, they shared the following
characteristics:
168
8.3. EXPERIMENT: VISUALIZATION OF MOODS
Figure 8.15: Facial expressions with their corresponding PAD intensities for docile
Mood: Exuberant
169
CHAPTER 8. EVALUATION
Figure 8.16: Mean values of the analysis of facial expressions of mood exuberant
Table 8.15: Mean values and Standard Deviations for expressions of Exuberant
The former results constrain our analysis to these expressions, which were further stud-
ied by obtaining the recognition hit rates for each dimension. Table 8.16 shows the hit
rates for the subset of the selected images.
The hit rates analysis shows that expressions with low pleasure were discarded, leaving
those expressions with medium to high pleasure and any level of arousal and dominance.
From these results we concluded that the activated AUs are AU 6, AU 12 for medium to
high positive pleasure values, and AU 5, AU 25, AU 26 for arousal with any intensity level.
170
8.3. EXPERIMENT: VISUALIZATION OF MOODS
It is also interesting to note that expressions combining low +P, high +A and high,
medium and low +D were identified as hostile mood. A possible explanation for these
results is that in the area close to low pleasure (positive and negative), the AU 12 responsible
for pulling the corner lips is not active. This in combination with open eyes (feature of
arousal) provokes a different effect than exuberance. Figure 8.17 shows a subset of the
expressions corresponding to each image of these values.
Figure 8.17: Facial expressions with their corresponding PAD intensities for exuberant
171
CHAPTER 8. EVALUATION
Mood: Hostile
This mood corresponds to the -P+A+D quadrant of the PAD model. Figure 8.18 presents
the distribution of the mean recognition values of each of the 27 images corresponding to
the mood hostile.
Figure 8.18: Mean values of the analysis of facial expressions of mood hostile
Table 8.17 presents the mean and standard deviation values of the expressions which
dimensions were identified as -P+A+D.
Table 8.17: Mean values and Standard Deviations for expressions of Hostile
172
8.3. EXPERIMENT: VISUALIZATION OF MOODS
The previous results allowed us to constrain our analysis to those expressions in which
subjects correctly recognized the -P+A+D dimensions. These let us with the 59.2% of the
expressions, for which the recognition hit rates were obtained. Table 8.18 shows the hit
rates for the subset of the selected images.
From the previous analysis it is shown that the main issue with these expressions was
the Dominance dimension, which presented very high standard deviations. Nevertheless, a
considerable number of expressions were assessed as -P+A+D, from which we extract the
following characteristics:
- Negative pleasure has any intensity level
- Positive arousal has also any intensity level
- Positive dominance is mainly assessed with intensity levels between medium and high
These results allow us to identify the AUs that are activated and describe the expres-
sions that can be associated to the mood hostile. The AUs are AU 4, AU 10, AU 15 for
medium to high positive dominance values, and AU 5, AU 25, AU 26 for arousal with any
intensity level.
173
CHAPTER 8. EVALUATION
Figure 8.19: Facial expressions with their corresponding PAD intensities for hostile
Mood: Relaxed
This mood corresponds to the +P-A+D quadrant of the PAD model. Therefore, we
expect expressions corresponding to this mood to be recognized as with positive pleasure
and dominance, and negative arousal.
Figure 8.20 presents the distribution of the mean recognition values of each of the 27
images corresponding to the mood relaxed.
Table 8.19 presents the mean and standard deviation values of the expressions identified
in +P-A+D.
Image Pleasure Arousal Dominance
Mean SD Mean SD Mean SD
04-Re-HMH 0.71 0.26 -0.07 0.67 0.14 0.55
05-Re-HMM 0.64 0.24 -0.43 0.34 0.21 0.48
06-Re-HML 0.43 0.45 -0.50 0.40 0.29 0.48
10-Re-MHH 0.64 0.23 -0.18 0.51 0.05 0.47
25-Re-LLH 0.25 0.26 -0.19 0.25 0.13 0.35
26-Re-LLM 0.25 0.27 -0.17 0.25 0.08 0.49
27-Re-LLL 0.13 0.22 -0.04 0.25 0.04 0.39
Table 8.19: Mean values and Standard Deviations for expressions of Relaxed
The previous results allowed us to constrain our analysis to those expressions in which
174
8.3. EXPERIMENT: VISUALIZATION OF MOODS
Figure 8.20: Mean values of the analysis of facial expressions of mood relaxed
subjects correctly recognized the +P-A+D dimensions. Therefore, we constrain the anal-
ysis to assess the hit rates for these two images. Table 8.20 shows the hit rates for the
subset of the selected images.
Image +P -A +D Number of Subjects
04-Re-HMH 7 3 3 7
05-Re-HMM 7 5 3 7
06-Re-HML 6 5 4 7
10-Re-MHH 11 4 3 11
From the hit rates analysis we observed that a reduced number of expressions were
indeed assessed as +P-A+D. The main issue was to evaluate the positive dominance,
which in all the 27 expressions obtained mean values close to zero, meaning that to the
question How dominant is Alfred? the majority of the subjects answered i dont know.
An example of this situation is the expressions tagged as 10-Re-MHH, which pleasure
dimension obtained a 100% of recognition rates, but arousal and dominance less than 36%.
The expressions that in the end were associated to the relaxed mood shared the following
characteristics:
- Positive pleasure with high intensity values
- Negative arousal with medium intensity values
- Positive dominance was independent of the other values
These results allow us to identify the AUs that are activated and describe the expres-
175
CHAPTER 8. EVALUATION
sions that can be associated to the mood relaxed. The AUs are AU 6, AU 12 for high positive
pleasure values, and AU 43 for arousal with medium intensity.
Figure 8.21 presents some of the expressions associated with the mood relaxed.
Figure 8.21: Facial expressions with their corresponding PAD intensities for mood relaxed
8.3.5 Discussion
The experiment that has been presented answered the hypothesis formulated at the begin-
ning of this section, which intended to prove that all moods have associated expressions
which are described by the AUs activated in their corresponding pleasure, arousal and
dominance dimension.
After the analysis of the results, we found that indeed the expressions associated with
a mood are described by the AUs in the octant of that mood. Given that this was an
exploratory experiment to see if the functions to locate AUs in the PAD space were valid,
we did not establish any percentage of error or success. Therefore, we used mean values and
standard deviations to compute the average perception of each expression. From this, we
obtained subsets of expressions which should be further evaluated to exhaustively validate
them.
Regarding the easiness of recognition of each dimension, the results concluded that:
- Pleasure was correctly identified in the vast majority of the cases, except for those in
which its values were close to zero, and it is seen as the dimension that gives meaning to
the expression.
- Arousal was also correctly identified in most of the cases, except on few cases when arousal
was negative and most of the subjects did not know how to assess it.
- Dominance presented most of the perception problems. Subjects were not sure how to
176
8.3. EXPERIMENT: VISUALIZATION OF MOODS
Mood (AUs)
Exuberant AU 6, AU 5, AU 12, AU 25, AU 26
Bored AU 1, AU 2, AU 4, AU 15, AU 43
Docile AU 1, AU 2, AU 12, AU 43
Hostile AU 4, AU 10, AU 5, AU 15, AU 25, AU 26
Anxious AU 1, AU 2, AU 4, AU 5, AU 15, AU 25, AU 26
Relaxed AU 6, AU 12, AU 43
Dependent AU 1, AU 2, AU 5, AU 12, AU 25, AU 26
Disdainful AU 4, AU 15, AU 43
177
CHAPTER 8. EVALUATION
One of the objectives of our work was the representation of context in order to generate
believable emotions in the character. Moreover, from previous experiments we observed
that a character can have the same facial expression for different emotions, and the inter-
pretation of these expressions may fail if we do not know enough about that character, or
the context in which that expression is occurring. Therefore, it is important not only to
focus on creating believable expressions but also to associate some beliefs (given by the
context) to those expressions.
8.4.1 Hypothesis
Given that exists a real facial expression for C (e.g. of an actor or an ordinary
person), the facial expression of the character is similar to the real one under C.
8.4.2 Methodology
To validate the former hypothesis we used the following methodology to perform the ex-
periment:
1. Selection and definition of scenes of a movie into the Context Representation Module.
It will give as output a set of emotions, from which the one with the greatest value
is chosen. If two emotions are elicited with the same intensity, we mix them using
the algorithm to generate intermediate expressions.
178
8.4. EXPERIMENT: CONTEXT REPRESENTATION
For this experiment we chose two movies: Leon (1994, directed by Luc Besson) and Down-
fall (Der Untergang) (2004, directed by Oliver Hirschbiegel). From each movie we selected
three scenes where it could be clearly seen the facial expressions of the real actors.
In Movie 1- Leon we selected the part where the main character Mathilda, a 12 years-
old girl enters her building, and finds out that her family has been brutally murdered. In
Movie 2- Downfall we selected the part when Hitler, already tired and defeated wants
to counterattack and his soldiers try to make him change his mind. In this last scenes, we
focused on one of the soldiers and his emotional reactions.
From the scenes of Movie 1 and Movie 2, we set up the events, characters preferences,
descriptions of the locations, admiration among characters, in order to be introduced in
the Context Representation module, and elicit the corresponding emotions. Tables 8.22
and 8.23 contains the elements used for the context representation.
Character: Mathilda
Goals: - Stay alive
Preferences: Gangsters (ST RON GLY BAD = 1.0)
Familys apartment (IN DIF F EREN T = 0.0)
Agent Admiration: Little brother (P OSIT IV E = 1.0)
Event Role of Character Emotions elicited
(Event 1) Mathilda comes home and notices a Mathilda executes the event fear = 0.4
gangster in front of the familys apartment. (ByMe)
She acts as if nothing happens
(N OT SAT ISF ACT ORY = 0.7)
(Event 2) Mathilda hears that her little brother The brother receives the event pity = 1.0
has been killed (OnOther)
(N OT SAT ISF ACT ORY = 1.0)
(Event 3) Mathilda rings the doorbell of Leon and Mathils receives the event sadness = 1.0
waits while the gangster is outside (OnMe)
(N OT SAT ISF ACT ORY = 1.0)
The emotions presented in the former tables, generated by each event, were the ones
that corresponded to the script of the movie annotated in a previous work that used the
same set of scenarios [138]. This correspondence between the emotion elicited and the
emotion in the script is a hint that the rules of the Semantic Model produce the right
emotions to the events.
Using the emotions elicited, we generated the facial expressions of the characters. For
179
CHAPTER 8. EVALUATION
movie Leon we used the character Alice, which resembled Mathilda. For the movie Downfall
we used Alfred, who quite resembled The Officer. Table 8.24 shows the virtual characters
and the actors in the events described in Tables 8.22 and 8.23.
These images were evaluated through an online survey with the intention to find
(1) if the facial expression of the virtual actor is believable in correspondence with the
event/scene, and (2) if the expressions of the real actor and the virtual character were
perceived as similar by the subjects.
180
(a) Leon - Event (1): fear (b) Leon - Event (2): pity (c) Leon - Event (3): sadness
8.4. EXPERIMENT: CONTEXT REPRESENTATION
(d) Downfall - Event (1): fear (e) Downfall - Event (2): discourage (f) Downfall - Event (3): disappointment
Table 8.24: Virtual and real actors from the movies Leon and Downfall
181
CHAPTER 8. EVALUATION
Questionnaire
61 subjects (30 male and 31 female) between 18 and 55 years old (MEAN = 28, SD =
5.45) participated in the experiment through an online questionnaire.
Before evaluating the scenes of each movie, we gave a synopsis of what was happening to
center the subject in the context. For the Movie 1- Leon the description was: Mathilda,
a 12-year old girl, comes home from the grocery. When she finds that her family has been
brutally murdered she seeks help from Leon, whose apartment is down the hall.
For the Movie 2- Downfall the description was: During the last days of the II WW,
Hitler is sick and exhausted; the defeat of the Third Reich is imminent. Hitler and his
officers are having a meeting in which he is planning impossible counterattacks and giving
orders to men who are already dead. His officers have serious doubts about their
leader but do not dare to speak up because of the reprimands..
Then, for each event showed in Tables 8.22 and 8.23, we asked the following questions:
- (Question 1) Facial expression elicited by context: The following image is the facial
expression of <nombre character> after the event described above. Do you agree this
would be the facial expression?
- (Question 2) Similarity between actors: Are the avatars expression (left) and the actors
expression (right) similar to each other?
For (Question 1) we used the images of Alice and Alfred shown in Table 8.24. For
(Question 2) we used the pair images Alice-actress and Alfred-actor shown in the same
table.
Each question was rated using a 5-item Likert Scale, where 1 corresponded to the
minimum degree (totally disagree or very different) and 5 to the maximum (totally agree or
very similar ). Figure 8.22 shows a page of the online questionnaire. The full questionnaire
can be found in the following URL: http://dmi.uib.es/ ugiv/diana/contextmodel/.
182
8.4. EXPERIMENT: CONTEXT REPRESENTATION
8.4.4 Results
The results obtained allow us to validate the proposed Context Representation Module, as
well as the generation of facial expressions in two avatars, one animated using MPEG-4
(Alice - Leon) and the other FACS (Alfred - Downfall ).
Table 8.25 presents the results for each question of each of the events taken from
Tables 8.22 and 8.23.
183
CHAPTER 8. EVALUATION
Table 8.25: Virtual and real actors from the movies Leon and Downfall
To analyze Question (1), we use the number of hits per item of the Likert-scale. We
observe that over 50% of the subjects agreed that the three expressions of Alice in Leon
corresponded to the events that generate them.
As for Alfred in the Downfall, the only expression for which 70.5% of the subjects
agreed that it was according to the scene was the one in Event (3) (Table 8.24), which
was the one with greatest emotional load (the emotion disappointment had an intensity of
0.9). In the Event (1) subjects disagreed with the facial expression of the avatar, which
indicates that the expression did not transmit the emotion of the scene. In the Event (2)
40.9% of the subjects agreed that the expression matched the situation, but 23% of the
subjects could not decide.
In the analysis of Question (2), subjects considered that the avatars were similar
and very similar to the actors in the Event (2) of the movie Leon and in the Event (3)
of the movie Downfall. In the remaining events, opinions of the participants were divided
obtaining almost equally results (less than 50%) for agreement and disagreement between
real and virtual expressions.
The reason why we did not use the mean values for the analysis is that they were
+3.0 with a standard deviation of +1.0 approximately. Therefore, we could not get any
conclusion based only on these values.
184
8.4. EXPERIMENT: CONTEXT REPRESENTATION
8.4.5 Discussion
Two hypothesis were formulated at the beginning of this section: (1) the generated facial
expressions of the avatar are in correspondence to the context that is represented; (2) there
is a similarity between the facial expressions generated in the avatar and those of the actors
when simulating the same context.
As for the first hypothesis we conclude that the facial expressions transmitted the
emotional content of the context, validating the output of our Context Representation
module. The cases where it failed corresponded to the expressions of Alfred, which were
not expressive enough due to the low intensity of the emotions in context.
As for the second hypothesis, although there were agreements of similarity, subjects
also focused in features as sweating (in the movie Downfall the actor is sweating), shape of
the mouth, and tears (in the case of Leon). Nevertheless, these observations gave us even
more hints about what participants take into consideration when perceiving an emotion in
the character even when context is given.
185
CHAPTER 8. EVALUATION
186
Chapter 9
Conclusions
The work that has been presented is the answer to those questions formulated at the
beginning of this thesis with the aim of developing a contextual and affective framework
that allows the creation of virtual characters capable of expressing emotions, mood and
personality through facial expressions and visual cues.
Therefore, this thesis is the result of an exhaustive research on different topics that
contribute to the representation and generation of context, as well as to the simulation
and visualization of affect. Among these topics are psychological theories of emotions,
mood and personality; computational models for the creation of interactive virtual charac-
ters; semantic web and ontologies; and visualization techniques for the generation of facial
expressions.
We begin this chapter by briefly reviewing each of the previous chapters, in order to
depict a general picture of the contributions and novelties presented in this thesis. Then,
a discussion section will present a more detailed view of the aspects that were considered
during this research, the results of the multiple experiments and evaluations which helped
to prove stated hypothesis and shed light on new theories, and finally the tasks to perform
as future work.
187
CHAPTER 9. CONCLUSIONS
9.1 Summary
This thesis started as a quest for the elements that are necessary to create believable
characters. Nevertheless, a large number of previous works have pursued this quest as
well, obtaining as a result computational models based on psychological theories which can
simulate up to a point believable characters.
But as the human being is complex enough, there are several elements which also have
to be considered to enrich the believability of these characters. One of this elements is the
context that surrounds the character, not only from a physical point of view, but also from
an affective point of view. One needs to think of the different things that could happen to
the characters, of all the things that they could do, what they like or dislike, what they
aim for, an so on. In other words, what the world makes the character feel.
Another element that has been not so extensively researched is the expression of moods
and personality. Thus we aimed to find a way to manifest these two affective elements in
the face of the character, and use them to enhance the expressions given by emotions,
making the character more believable.
Therefore, knowing which were the research objectives and the results we wanted to
obtain, we started a deep investigation of psychological theories that define emotions,
moods and personality (Chapter 2). Also, we researched on the different computational
models that have been developed to simulate virtual characters (Chapter 3). From this state
of art, we found that the majority of models share the same psychological background. It
means, they use the OCC model to represent emotions, the Five-Factor Model to represent
personality, and just until recently a few works have used the PAD Space to represent
moods. On the other hand, the techniques implemented to simulate the behaviors of the
characters goes from neural networks, to mathematical-logical functions or scripting.
However, the answer to our questions remained unsolved: how to represent the context
of the character from a physical and affective point of view, and how to visualize the mood
and personality. That is why we proposed a novel framework that comprises a context
representation, affect processing and visualization of these affective traits (Chapter 4).
The context representation was achieved through the design and implementation of a
Semantic Model, that making use of ontologies allowed us to define all those concepts that
are important to model the internal and external world of the characters (Chapter 5). The
novelty of our ontologies is that they do not just represent the concepts of the physical world
and affective traits of the character, but allows inference on them, obtaining emotional
188
9.2. DISCUSSION
9.2 Discussion
The contextual and affective framework proposed in this thesis required an exhaustive
analysis and study of different fields that have been combined in order to create a complete
process of character generation. That is why through the different chapters psychology,
ontologies, affective computing, computer graphics and facial annotation standards are
interlaced and treated to a depth that is sufficient to achieve the objectives of this thesis.
In the case of ontologies, we designed and implemented two ontologies that took into
account external and internal aspects of the character and its virtual world. Our intention
was not to provide new semantics and formal validation for them. Our real aim was to
provide a tool that allows the definition of meta-concepts related to the character and its
world, make inferences about them, and therefore, easily generate events that occur in that
world and automatically obtain their influence on the characters emotional state. In this
way, a storyteller could just define the events of the world and the psychological profile of
189
CHAPTER 9. CONCLUSIONS
the character, and then stories would be simulated by only re-using and combining this
information.
Regarding affective computing and psychology, our intention was not to create a new
paradigm or a new computational model for the creation of virtual characters. Our true
objective was to use previous works and theories and combined them in a new way that
permits to build a bridge between the contextual representation and the visual representa-
tion of affect. That is way we made use of the well-known OCC model, Five Factor model
and PAD Space for generation of emotions, personality and mood, respectively; as well as
of the ALMA model to get the basis of a system that simulated the interaction of these
affective traits.
Finally, this thesis did not deal with new computer graphics techniques, instead we
used low-cost and efficient facial animation and facial actions description standards like
the MPEG-4 and FACS to implement algorithms for the generation of novel and original
parameterizable facial expressions, like the expressions of all the intermediate emotions
of the OCC. Moreover, due to the advantages of parameterization, these algorithms are
independent of animation procedures, being extendible to other standards and visualiza-
tion techniques. Another novelty in this work is the formulation and implementation of
equations that allowed us to create the expressions for a very important affective trait, as
it is the mood.
This thesis is also strongly grounded on experimentation and evaluation. The several
hypothesis stated for visualization of affect and generation of context led to experiments
that not only validated our results, but assessed the perception on the subjects side,
giving hints and opening new ways to continue researching in topics that could enormously
enhance the believability of characters.
In the case of the evaluation of expressions of emotions, the results demonstrated that
subjects could recognized the expressions used for the generation of the expressions of
intermediate emotions. Even more, in some cases, subjects rated the obtained intermediate
expression with the proper emotional term. Besides suggestions related to the morphology
of the character (e.g. hair, wrinkles), the numbers showed that expressions obtained in
general highly recognition scores.
As for the moods, we performed a questionnaire where subjects could assess expressions
of an avatar according to its pleasure (P), arousal (A) and dominance (D) intensity values.
The results of this test conveyed two important conclusions. The first one is that all
the 8 moods defined by the PAD model have their corresponding facial expressions. Some
190
9.2. DISCUSSION
moods like exuberant, bored or hostile presented a large number of expressions, while other
moods like docile or dependent presented one or two. This difference was mainly due to the
difficulty of assessment of Dominance in static images. The second important conclusion
is that the functions that were formulated to map AUs into the PAD Space were valid.
These results are of great significance because knowing the AUs that describe the
expressions of certain mood, other emotional facial expressions can be enhanced and truly
reflect the affective state of the character.
Another aspect that was studied from a purely experimental approach was the percep-
tion of personality traits. In this thesis we presented the initial steps of a research that
can be easily branched out, because personality is a component that is always expressed
but difficult to simulate in a procedural way.
To evaluate the perception of personality we produced a set of configurations for head
position and eye gaze, in a way that they do not affect the facial features but somehow
enhance the emotional expression. The results were very encouraging because subjects
indeed associate extraversion, agreeableness and neuroticism to some configurations. For
the Alfred character, the upwards-sideways head orientation was related to extraver-
sion, downwards-center head orientation to agreeableness, and center-sideways head
orientation to emotional stability. However, eye gaze did not influence the perception of
personality traits.
All these previous findings represent a great advance in the area of visualization of
affective traits, because we have more information about all the elements that should be
considered when making a character look more believable from an affective point of view.
Finally, a very original experiment was the one to validate the Context Representation
module. The idea was to see if subjects could better assessed a facial expression if the
context of that expression was given.
The ground data we used to validate the module were the scenes of two movies, so we
would have real facial expressions of real actors to compare with when the same context
is elicited. The movies were Leon and Downfall, basically because they were scripted in
a way we could easily use in our module, and second because the characters Mathilda in
Leon and The soldier in the movie Downfall, resembled our characters Alice and Alfred,
respectively.
Once again, the results proved that the ontologies we defined and the contextual module
we implemented has the potential to feasible reproduce certain context and elicit suitable
emotions. For the movie Leon, the participants could recognize in the expressions of Alice
191
CHAPTER 9. CONCLUSIONS
the emotions triggered by the events. For the movie Downfall, the results were not that
satisfactory, mainly because of the repressed and passive emotional content of the chosen
scenes, and the lack of notorious expressions in the actor, and in the character.
In summary, the framework we propose and the experiment results we obtained can
be applied in different applications where it is required to have a story or the interaction
between the character and its environment. For instance, educative applications, where the
student needs to learn about certain topics with the help of a virtual tutor or simulations
in a virtual world; story simulation (kind of storytelling) for representing movies; or theme-
oriented situations where the story/script is basically generated by the interaction of the
character and its world.
192
9.4. PUBLICATIONS AND CONTRIBUTIONS
9.4.1 Journals
D. Arellano, J. Varona, F J. Perales. Why do I feel like this? - The importance
of context representation for emotion elicitation. International Journal of Synthetic
Emotions (IJSE), 2(2), 2011. DOI: 10.4018/IJSE. ISSN: 1947-9093.
193
CHAPTER 9. CONCLUSIONS
9.4.2 Proceedings
D. Arellano, N. Bee, K. Janowski, E. Andre, J. Varona, F. J. Perales. Influence of
Head Orientation in Perception of Personality Traits in Virtual Agents. AAMAS
2011 (Short paper), Taiwan, May, 2011.
I. Lera, D. Arellano, J. Varona, C. Juiz and R. Puigjaner. Semantic Model for Facial
Emotion to improve the human computer interaction in AML. In UCAmI 2008, Vol.
51/2009, Salamanca, Spain, pp. 139-148, 2008.
9.4.3 Workshops
D. Arellano, I. Lera, J. Varona and F. J. Perales. Generating Affective Characters for
Assistive Applications. EMOTIONS & MACHINES Workshop, Geneva, Switzerland,
August 2009.
194
Bibliography
[1] I. Albrecht, M. Schroder, J. Haber, and H. Seidel. Mixed feelings: Expression of non-
basic emotions in a muscle-based talking head. Journal of Virtual Reality. Special
issue on Language, Speech and Gesture for Virtual Reality, 8(4):201212, 2005.
[2] R. D. Aldrich. What ever happened to baby jane? [film]. Film, 1962. Warner Bros.
[7] A. Arya, L. N. Jefferies, J. T. Enns, and S. DiPaola. Facial actions as visual cues
for personality. Journal of Visualization and Computer Animation, 17(3-4):371382,
2006.
[8] K. Balci. XfaceEd: Authoring tool for embodied conversational agents. In ICMI,
pages 208213, 2005.
195
BIBLIOGRAPHY
196
BIBLIOGRAPHY
197
BIBLIOGRAPHY
[29] T. Cochrane. 8 dimensions for the emotions. Social Science Information. Special
issue The language of emotion: conceptual and cultural issues, 48:379420, 2009.
[30] M. Courgeon, S. Buisine, and J.-C. Martin. Impact of expressive wrinkles on per-
ception of a virtual characters facial expressions of emotions. Proceedings of the 9th
International Conference on Intelligent Virtual Agents, pages 201214, 2009.
[31] M. Courgeon, C. Clavel, and J.-C. Martin. Appraising emotional events during a real-
time interactive game. In Proceedings of the International Workshop on Affective-
Aware Virtual Agents and Social Robots, AFFINE 09, pages 7:17:5, New York, NY,
USA, 2009. ACM.
[32] R. Cowie, E. DouglasCowie, S. Savvidou, E. McMahon, M. Sawey, and M. Schroder.
Feeltrace: An instrument for recording perceived emotion in real time. In Proceedings
of the ISCA Workshop on Speech and Emotion, pages 1924, 2000.
[33] C. Darwin. The Expression of Emotions in Man and Animals. London: Murray,
1873.
[34] B. De Carolis, V. Carofiglio, and C. Pelachaud. From discourse plans to believ-
able behaviour generation. In Proceedings of the International National Language
Generation Conference, New York, 2002.
[35] F. de Rosis, C. Pelachaud, I. Poggi, V. Carofiglio, and B. D. Carolis. From gretas
mind to her face: modelling the dynamics of affective states in a conversational
embodied agent. International Journal of Human-Computer Studies, 59(1-2):81118,
2003.
[36] A. K. Dey, G. D. Abowd, and D. Salber. A conceptual framework and a toolkit for
supporting the rapid prototyping of context-aware applications. Human-Computer
Interaction, 16:97166, December 2001.
[37] J. Dias. Fearnot!: Creating emotional autonomous synthetic characters for empathic
interactions. Masters thesis, Universidade Tecnica de Lisboa, Instituto Superior
Tecnico, 2005.
[38] J. Dias and A. Paiva. Feeling and reasoning: a computational model for emo-
tional agents. In Proceedings of 12th Portuguese Conference on Artificial Intelligence,
EPIA, pages 127140, 2005.
[39] P. Doyle. Believability through context using knowledge in the world to create
intelligent characters. In Proceedings of the First International Joint Conference on
198
BIBLIOGRAPHY
Autonomous Agents and Multiagent Systems: part 1, AAMAS 02, pages 342349,
New York, NY, USA, 2002. ACM.
[40] L. Edelstein. The Writers Guide to Character Traits. Writers Digest Books, 2
edition, 2009.
[41] A. Egges, S. Kshirsagar, and N. Magnenat-Thalmann. Generic personality and emo-
tion simulation for conversational agents: Research articles. Comput. Animat. Virtual
Worlds, 15(1):113, 2004.
[42] P. Ekman. Emotion in the Human Face. Cambridge University Press, 1982.
[43] P. Ekman. Moods, emotions, traits. In P. E. . R. Davidson, editor, The nature of
emotion, pages 5658. New York: Oxford University Press, 1994.
[44] P. Ekman. Basic emotions. In T. Dalgleish and M. Power, editors, Handbook of
Cognition and Emotion, chapter 3. Sussex, U.K.: John Wiley & Sons, Ltd., 1999.
[45] P. Ekman. Facial expressions. In T. Dalgleish and M. Power, editors, Handbook of
Cognition and Emotion, chapter 16. New York: John Wiley & Sons Ltd, 1999.
[46] P. Ekman. Darwins contributions to our understanding of emotional expressions.
Philosophical Transactions of the Royal Society Biological Sciences, 364(1535):3449
3451, 2009.
[47] P. Ekman and W. Friesen. Unmasking the Face. New Jersey: Prentice-Hall, 1975.
[48] P. Ekman, W. Friesen, and J. Hager. The Facial Action Coding System. Weidenfeld
& Nicolson, London, UK, 2nd edition, 2002.
[49] C. Elliott. I picked up catapia and other stories: a multimodal approach to expres-
sivity for emotionally intelligent agents. In Proceedings of the first international
conference on Autonomous agents, AGENTS 97, pages 451457, New York, NY,
USA, 1997. ACM.
[50] C. Elliott. Hunting for the holy grail with emotionally intelligent virtual actors.
SIGART Bull., 9(1):2028, 1998.
[51] H. J. Eysenck and S. B. G. Eysenck. Manual of the Eysenck Personality Question-
naire. London: Hodder and Stoughton, 1975.
[52] S. Eysenck. Citation classic. the measurement of psychoticism: a study of factor
stability and reliability. CC/SOC BEHAV SCI, 1(35):12, 1986.
199
BIBLIOGRAPHY
[53] M. Fabri, D. J. Moore, and D. J. Hobbs. Designing avatars for social interactions. In
L. Canamero and R. A. (eds.), editors, Animating Expressive Characters for Social
Interaction, Advances in Consciousness Research Series. Benjamins Publishing, 2008.
[54] G. Faigin. The Artists Complete Guide to Facial Expressions. Watson-Guptill, New
York, 1990.
[55] C. Felbaum. Wordnet, an Electronic Lexical Database for English. Cambridge: MIT
Press, 1998.
[56] E. Figa and P. Tarau. The vista architecture: experiencing stories through virtual
storytelling agents. SIGGROUP Bull., 23(2):2728, 2002.
[58] R. Forchheimer and I. S. Pandzic. MPEG-4 facial animation The Standard, Imple-
mentation and Applications. John Wiley and Sons, West Sussex, UK, 1996.
[59] N. H. Frijda. The Emotions. Cambridge ; New York : Cambridge University Press;
Paris: Editions de la Maison des sciences de lhomme, 1986.
[60] N. H. Frijda. The laws of emotion. In . N. S. Jennifer M. Jenkins, Keith Oatley, editor,
Human Emotions: A Reader, pages 271287. Malden, MA: Blackwell Publishers,
1998.
[63] P. Gebhard. ALMA: a layered model of affect. In AAMAS 05: Proceedings of the
fourth international joint conference on Autonomous agents and multiagent systems,
pages 2936, NY, USA, 2005. ACM.
200
BIBLIOGRAPHY
[65] P. Gebhard and K. H. Kipp. Are computer-generated emotions and moods plausible
to humans?. In IVA, volume 4133 of Lecture Notes in Computer Science, pages
343356. Springer, 2006.
[66] S. D. Gosling, P. J. Rentfrow, and W. B. S. Jr. A very brief measure of the big-five
personality domains. Journal of Research in Personality, 37(6):504528, 2003.
[67] K. Grammer and E. Oberzaucher. The reconstruction of facial expressions in em-
bodied systems: New approaches to an old problem. ZiF Mitteilungen, 2:1431,
2006.
[68] W. O. W. Group. Owl 2 web ontology language. document overview. w3c recommen-
dation 27 october 2009. Latest version available at http://www.w3.org/TR/owl2-
overview/. Last reviewed: 09/19/2010.
[69] M. Gruninger and M. Fox. Methodology for the design and evaluation of ontologies.
In IJCAI95, Workshop on Basic Ontological Issues in Knowledge Sharing, 1995.
[70] M. Gutierrez, A. Garca-Rojas, D. Thalmann, F. Vexo, L. Moccozet, N. Magnenat-
Thalmann, M. Mortara, and M. Spagnuolo. An ontology of virtual humans: Incor-
porating semantics into human shapes. Vis. Comput., 23(3):207218, 2007.
[71] D. Heckmann, T. Schwartz, B. Brandherm, M. Schmitz, and M. von Wilamowitz-
Moellendorff. Gumo: The general user model ontology. User Modeling 2005, pages
428432, 2005.
[72] L. Hewlett-Packard Development Company. Jena: A Semantic Web Framework for
Java, 2008. version 2.5.6. http://jena.sourceforge.net/.
[73] J. R. Hobbs and F. Pan. Time ontology in owl. Latest version available at
http://www.w3.org/TR/owl-time/. Last reviewed: 25/06/2011.
[74] W. K. Hofstee, B. de Raad, and L. R. Goldberg. Integration of the big five and cir-
cumplex approaches to trait structure. Journal of Personality and Social Psychology,
63(1):146163, 1992.
[75] D. C. Howe. Rita.wordnet. http://www.rednoise.org/rita/wordnet/documentation/index.htm.
Last reviewed: 08052011.
[76] S. Inversions. http://www.facegen.com/modeller.htm.
[77] K. Isbister and C. Nass. Consistency of personality in interactive characters: ver-
bal cues, non-verbal cues, and user characteristics. Int. J. Hum.-Comput. Stud.,
53(2):251267, 2000.
201
BIBLIOGRAPHY
[78] S. Kaiser and T. Wehrle. Facial expressions in social interactions: Beyond basic
emotions, volume Advances in Consciousness Research (Vol.74): Animating Expres-
sive Characters for Social Interactions, chapter 4, pages 5369. Amsterdam : John
Benjamins Publishing Company, 2008.
[79] Z. Kasap and N. Magnenat-Thalmann. Intelligent virtual humans with autonomy and
personality: State-of-the-art. Intelligent Decision Technologies, 1(1-2):315, 2007.
[86] B. Lance and S. Marsella. Glances, glares, and glowering: how should a virtual
human express emotion through gaze? Autonomous Agents and Multi-Agent Systems,
20(1):5069, 2010.
202
BIBLIOGRAPHY
[88] Y. li Tian, T. Kanade, and J. F. Cohn. Recognizing action units for facial expression
analysis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 23:97
115, 2001.
[89] J. M. Lopez, R. Gil, R. Garca, I. Cearreta, and N. Garay. Towards an ontology
for describing emotions. In Proceedings of the 1st world summit on The Knowledge
Society: Emerging Technologies and Information Systems for the Knowledge Society,
WSKS 08, pages 96104, Berlin, Heidelberg, 2008. Springer-Verlag.
[90] C. Mackay, T. Cox, G. Burrows, and T. Lazzerini. An inventory for the measure-
ment of self-reported stress and arousal. The British Journal of Social and Clinical
Psychology, 17(3):283284, 1978.
[91] N. Magnenat-Thalmann and Z. Kasap. Modelling socially intelligent virtual
humans. In VRCAI 09: Proceedings of the 8th International Conference
on Virtual Reality Continuum and its Applications in Industry. ACM, 2009.
http://www.vrcai2009.com/socially intelligentVH.pdf.
[92] L. Malatesta, A. Raouzaiou, K. Karpouzis, and S. Kollias. Towards modeling em-
bodied conversational agent character profiles using appraisal theory predictions in
expression synthesis. Applied Intelligence, 30(1):5864, 2009.
[93] M. Mancini and C. Pelachaud. Implementing distinctive behavior for conversational
agents. Gesture-Based Human-Computer Interaction and Simulation, Lecture Notes
in Computer Science, 5085/2009:163174, 2009.
[94] S. Marsella and J. Gratch. Ema: A computational model of appraisal dy-
namics. In European Meeting on Cybernetics and Systems Research, 2006.
http://www.ict.usc.edu/ marsella/publications/N Emcsr Marsella.pdf.
[95] S. Marsella and J. Gratch. Ema: A process model of appraisal dynamics. Journal of
Cognitive Systems Research, 10(1):7090, 2009.
[96] G. Matthews, D. Jones, and A. Chamberlain. Refining the measurement of mood:
the uwist mood adjetive checklist. British Journal of Psychology, 81:1742, 1990.
[97] R. McCrae and O. John. An introduction to the five-factor model and its applications.
Journal of Personality, 60(1):175215, 1992.
[98] R. R. McCrae. The five-factor model of personality traits: consensus and contro-
versy. In . G. M. P. J. Corr, editor, Cambridge Handbook of Personality Psychology,
chapter 9, pages 148161. Oxford: Oxford University Press, 2009.
203
BIBLIOGRAPHY
204
BIBLIOGRAPHY
205
BIBLIOGRAPHY
206
BIBLIOGRAPHY
[133] J. Russell. Core affect and the psychological construction of emotion. Psychological
Review, 110(1):145172, 2003.
[137] M. Sagar. Facial performance capture and expressive translation for king kong. In
ACM SIGGRAPH 2006 Sketches, SIGGRAPH 06, New York, NY, USA, 2006. ACM.
[138] R. Schapp. Enhancing video game characters with emotion, mood and personality.
Masters thesis, Delft University of Technology, 2009.
[141] T. Strang and C. Linnhoff-Popien. A context modeling survey. In UbiComp 1st In-
ternational Workshop on Advanced Context Modelling, Reasoning and Management,
pages 3141, 2004.
[142] I. Swartjes and M. Theune. The virtual storyteller: story generation by simulation.
Twentieth Belgian-Netherlands Conference on Artificial Intelligence, BNAIC, 2008.
[143] B. Szekely and J. Betz. Jastor: Typesafe, ontology driven rdf access from java. Latest
version available at http://jastor.sourceforge.net/. Last reviewed: 25/06/2011.
[144] M. Tekalp and J. Ostermann. Face and 2-D mesh animation in MPEG-4. Signal
Processing: Image Communication, 15(4-5):387421, 2000.
207
BIBLIOGRAPHY
[146] Y. Tong, W. Liao, and Q. Ji. Inferring facial action units with causal relations. In
Proceedings of the 2006 IEEE Computer Society Conference on Computer Vision and
Pattern Recognition (CVPR06), 2006.
[149] X. H. Wang, D. Q. Zhang, T. Gu, and H. K. Pung. Ontology based context modeling
and reasoning using owl. In Proceedings of the Second IEEE Annual Conference on
Pervasive Computing and Communications Workshops, PERCOMW 04, pages 18,
Washington, DC, USA, 2004. IEEE Computer Society.
[151] D. Watson and L. A. Clark. On traits and temperament: general and specific factors
of emotional experience and their relation to the five-factor model. J Pers, 60(2):441
476, 1992.
[152] D. Watson and L. A. Clark. The panas-x: Manual for the pos-
itive and negative affect schedule - expanded form. Available from
http://www.psychology.uiowa.edu/Faculty/Watson/PANAS-X.pdf, 1994.
208
BIBLIOGRAPHY
209
BIBLIOGRAPHY
210
Appendix A
Ontology Rules
The following rules are specified for emotion elicitation in JENA format. Thus this is
the way they were defined in our application. The format is:
name-of-the-rule: a1 a2 ... an = c1 c2 ... cn ,
where ai and ci are predicates of the form: subject, relation, object.
A subject is in most of the cases a variable that is taken from the database. To use it in
the rule it is preceded by a question mark (?). For instance, the variable ?character refers
to the character to whom the rules are applied.
A relation is the link between the subject and the object. It is taken from the ontology
itself through an URI that indicates where it has defined. In our case the URIs that has
been defined are:
- personalityOntology: http://dmi.uib.es/ ugiv/ontologies/personalityEmotion
- event: http://dmi.uib.es/ ugiv/ontologies/event
- rdf-syntax: http://www.w3.org/1999/02/22-rdf-syntax-ns (this is the definition of
RDF concepts according to the standard)
211
APPENDIX A. ONTOLOGY RULES
Rule: NotSatisfactionByMeEmotions
If the event is NotSatisfactory and the role is ByMe, then the rules are:
<EventRelation><hasSatisfactionValue><EventSatisfactionScale>
?eventSatScale <rdf-syntax#type><personalityOntology#NotSatisfactory>
?character <personalityOntology#hasEventSatisfaction> ?eventRelation
?character <event#hasRole> ?agentRole
?agentRole <rdf-syntax#type> <event#ByMe>
=
?agentRole <personalityEmotion#hasNegativeSelfEmotion><personalityEmotion#Shame>
?agentRole <personalityEmotion#hasNegativeSelfEmotion><personalityEmotion#Remorse>
Rule: SatisfactionByMeEmotions
If the event is Satisfactory and the role is ByMe, then the rules are:
?eventRelation <personalityEmotion#hasSatisfactionValue> ?eventSatScale
?eventSatScale <rdf-syntax#type> <personalityEmotion#Satisfactory>
?character <personalityEmotion#hasEventSatisfaction> ?eventRelation
?character <event#hasRole> ?agentRole
?agentRole <rdf-syntax#type><event#ByMe>
=
?agentRole <personalityEmotion#hasPositiveSelfEmotion><personalityEmotion#Pride>
?agentRole <personalityEmotion#hasPositiveSelfEmotion><personalityEmotion#Gratification>
Rule: IndifferentByMeEmotions
If the event does not have a level of satisfaction (i.e. it is Indifferent) and the role is ByMe, then the rules
are:
?eventRelation <personalityEmotion#hasSatisfactionValue> ?eventSatScale
?eventSatScale <rdf-syntax#type> <personalityEmotion#IndifferentE>
?character <personalityEmotion#hasEventSatisfaction> ?eventRelation
?character <event#hasRole> ?agentRole
?agentRole <rdf-syntax#type> <event#ByMe>
=
?agentRole <personalityEmotion#hasSelfEmotion><personalityEmotion#Neutral>
Rule: SatisfactionOnMeEmotionsGoal
If the event is Satisfactory, it means the realization of a goal, and the role is OnMe, then the rules are:
?eventRelation <personalityEmotion#hasEvent> ?event
?eventRelation <personalityEmotion#hasSatisfactionValue> ?satScale
?satScale <rdf-syntax#type> <personalityEmotion#Satisfactory>
?character <personalityEmotion#hasEventSatisfaction> ?eventRelation
?character <event#hasRole> ?agentRole
?agentRole <rdf-syntax#type> <event#OnMe>
?character <personalityEmotion#hasGoal>?event
=
?agentRole <personalityEmotion#hasPositiveSelfEmotion><personalityEmotion#Satisfaction>
?agentRole <personalityEmotion#hasPositiveSelfEmotion> <personalityEmotion#Relief>
Rule: SatisfactionOnMeEmotions
If the event is Satisfactory and the role is OnMe, then the rules are:
212
?eventRelation <personalityEmotion#hasEvent> ?event
?eventRelation <personalityEmotion#hasSatisfactionValue> ?satScale
?satScale <rdf-syntax#type><personalityEmotion#Satisfactory>
?character <personalityEmotion#hasEventSatisfaction> ?eventRelation
?character <event#hasRole> ?agentRole
?agentRole <rdf-syntax#type> <event#OnMe>
=
?agentRole <personalityEmotion#hasPositiveSelfEmotion> <personalityEmotion#Joy>
Rule: NotSatisfactionOnMeEmotionsGoal
If the event is NotSatisfactory, it is related to a goal, and the role is OnMe, then the rules are:
?eventRelation <personalityEmotion#hasEvent> ?event
?eventRelation <personalityEmotion#hasSatisfactionValue> ?satScale
?satScale <rdf-syntax#type> <personalityEmotion#NotSatisfactory>
?character <personalityEmotion#hasEventSatisfaction > ?eventRelation
?character <event#hasRole> ?agentRole
?agentRole <rdf-syntax#type> <event#OnMe>
?character <personalityEmotion#hasGoal> ?event
=
?agentRole <personalityEmotion#hasNegativeSelfEmotion> <personalityEmotion#Disappointment>
?agentRole <personalityEmotion#hasNegativeSelfEmotion> <personalityEmotion#Fear>
Rule: NotSatisfactionOnMeEmotions
If the event is NotSatisfactory and the role is OnMe, then the rules are:
?eventRelation <personalityEmotion#hasEvent> ?event
?eventRelation <personalityEmotion#hasSatisfactionValue> ?satScale
?satScale <rdf-syntax#type> <personalityEmotion#NotSatisfactory>
?character <personalityEmotion#hasEventSatisfaction> ?eventRelation
?character <event#hasRole> ?agentRole
?agentRole <rdf-syntax#type> <event#OnMe>
=
?agentRole <personalityEmotion#hasNegativeSelfEmotion> <personalityEmotion#Sadness>
Rule: PositiveAdmirationByOtherEmotions
The agent who has the role ByOther is positively admired by the agent who evaluates the event. Then the
rules are:
?agent <personalityEmotion#feels> ?admirationScale
?admirationScale <personalityEmotion#feelsFor> ?otherAgent
?admirationScale <rdf-syntax#type> <personalityEmotion#Positive>
?otherAgentRole <event#hasEmotionalAgent> ?otherAgent
?otherAgentRole <rdf-syntax#type><ontologies/event#ByOther>
=
?otherAgentRole <personalityEmotion#hasPositiveOtherEmotion> <personalityEmotion#Admiration>
?otherAgentRole <personalityEmotion#hasNegativeOtherEmotion><personalityEmotion#Reproach>
Rule: NegativeAdmirationByOtherEmotions
The agent who has the role ByOther is negatively admired by the agent who evaluates the event. Then the
rules are:
213
APPENDIX A. ONTOLOGY RULES
Rule: PositiveAdmirationOnOtherEmotions
The agent who has the role OnOther is positively admired by the agent who evaluates the event. Then the
rules are:
?agent <personalityEmotion#feels> ?admirationScale
?admirationScale <personalityEmotion#feelsFor> ?otherAgent
?admirationScale <rdf-syntax#type> <personalityEmotion#Positive>
?otherAgentRole <event#hasEmotionalAgent> ?otherAgent
?otherAgentRole <rdf-syntax#type> <event#OnOther>
=
?otherAgentRole <personalityEmotion#hasPositiveOtherEmotion> <personalityEmotion#Joy>
?otherAgentRole <personalityEmotion#hasNegativeOtherEmotion> <personalityEmotion#Pity>
Rule: NegativeAdmirationOnOtherEmotions
The agent who has the role OnOther is negatively admired by the agent who evaluates the event. Then the
rules are:
?agent <personalityEmotion#feels> ?admirationScale
?admirationScale <personalityEmotion#feelsFor> ?otherAgent
?admirationScale <rdf-syntax#type> <personalityEmotion#Negative>
?otherAgentRole <event#hasEmotionalAgent> ?otherAgent
?otherAgentRole <rdf-syntax#type> <event#OnOther>
=
?otherAgentRole <personalityEmotion#hasPositiveOtherEmotion> <personalityEmotion#Gloating>
?otherAgentRole <personalityEmotion#hasNegativeOtherEmotion> <personalityEmotion#Resentment>
Rule: EmotiveScaleSGEmotions
These are the emotions elicited due to a StronglyGood event. Then the rules are:
?emotiveScale <rdf-syntax#type> <personalityEmotion#StronglyGood>
=
?emotiveScale <personalityEmotion#hasPositiveEmotion> <personalityEmotion#Love>
?emotiveScale <personalityEmotion#hasEmotion><personalityEmotion#Love>
Rule: EmotiveScaleGEmotions
These are the emotions elicited due to a Good event. Then the rules are:
?emotiveScale <rdf-syntax#type> <personalityEmotion#Good>
=
?emotiveScale <personalityEmotion#hasPositiveEmotion> <personalityEmotion#Liking>
?emotiveScale <personalityEmotion#hasEmotion><personalityEmotion#Liking>
Rule: EmotiveScaleIEmotions
These are the emotions elicited due to a Indifferent event. Then the rules are:
?emotiveScale <rdf-syntax#type> <personalityEmotion#Indifferent>
=
214
?emotiveScale <personalityEmotion#hasEmotion><personalityEmotion#Neutral>
Rule: EmotiveScaleBEmotions
These are the emotions elicited due to a Bad event. Then the rules are:
?emotiveScale <rdf-syntax#type> <personalityEmotion#Bad>
=
?emotiveScale <personalityEmotion#hasNegativeEmotion> <personalityEmotion#Disliking>
?emotiveScale <personalityEmotion#hasEmotion><personalityEmotion#Disliking>
Rule: EmotiveScaleSBEmotions
These are the emotions elicited due to a StronglyBad event. Then the rules are:
?emotiveScale <rdf-syntax#type> <personalityEmotion#StronglyBad>
=
?emotiveScale <personalityEmotion#hasNegativeEmotion> <personalityEmotion#Hate>
?emotiveScale <personalityEmotion#hasNegativeEmotion> <personalityEmotion#Fear>
?emotiveScale <personalityEmotion#hasEmotion><personalityEmotion#Hate>
?emotiveScale <personalityEmotion#hasEmotion><personalityEmotion#Fear>
Rule: IsGoal
Allows to know if the event is a Goal. Then the rules are:
?agent <personalityEmotion#hasGoal> ?event
=
?event <personalityEmotion#isGoalOf> ?agent
215
APPENDIX A. ONTOLOGY RULES
216
Appendix B
B.1 Mapping
Table B.1 presents the mapping done from Facial Animation Parameters (FAPs) into Action
Units (AUs), based on the movements performed and the muscles involved.
Table B.1: Mapping of Facial Animation Parameters (FAPs) into Action Units (AUs)
217
APPENDIX B. MAPPING FROM FAPS TO AUS
218
B.1. MAPPING
219
APPENDIX B. MAPPING FROM FAPS TO AUS
220
B.1. MAPPING
221
APPENDIX B. MAPPING FROM FAPS TO AUS
1
This mapping is valid with FAPs with negative intensities - upper movement
2
Images for AUs: 1, 2, 4, 5, 6, 7, 9, 10, 11, 13, 15, 17, 18, 22, 25, 26, 27, 43b, 43d, 43, 44, 51, 52, 53, 54,
55, 56 were taken from: http://www.cs.cmu.edu/ face/facs.htm
3
Images for AUs: 12, 14, 16, 20, 23, 24, 28, 29, 46, 61, 62, 63, 64 were taken from:
http://micromovimiento.com/?p=933#more-933
4
AD instead of AU stands for Action Descriptor
222
B.2. OPPOSITE AUS
Table B.3 shows a set of AUs - 16, 25 and 26, which are opposed to AU17. As can
be seen in the table, AUs 16 and 25 describe a movement that involves FAPs 10 and 11
with and downward direction, while AU17 has a correspondence with these FAPs but with
upward direction. Regarding AU26, although it does not share similar FAPs with AU17,
the movement described by the former it is opposed to the one described by the latter.
Table B.4 shows two sets of AUs that are also opposed. On one hand, AUs 23 and
24 describe movements that involve a closure of the mouth, although they do not involve
the same set of FAPs. On the other hand, AUs 25, 26 and 27 describe movements that
imply openness of the mouth. Here we can see that AU24 is related to FAPs (5, 10 and
11) which are oriented upwards, while AU25 and AU27 are related to this set of FAPs but
downwards. Although AU26 is not related to the former FAPs, it describes a movement
that is opposed to the ones described by AUs 23 and 24.
223
APPENDIX B. MAPPING FROM FAPS TO AUS
Finally, Table B.5 presents AUs related with stretching and tightening of lips. AU20
is related to FAPs 6 and 7, which are oriented in a way that stretch the lips. Opposite to
that are AUs 22 and 23, which are also related to FAPs 6 and 7, but oriented in a way
that squash the lips (tightening or funneling them).
AUs Combinations
Table B.6 shows a set of combinations, collected from different sources that worked with
facial expression recognition. It gives us a hint of how AUs can be combined to create
different facial expressions.
On the other hand, Table B.7 shows a set of combinations that are not possible in
human faces.
224
B.2. OPPOSITE AUS
225
APPENDIX B. MAPPING FROM FAPS TO AUS
226
B.2. OPPOSITE AUS
AUs Source
15 + 25 [146]
12 + 15 own observation
16 + 17 comparison using related FAPs
17 + 26 comparison using related FAPs
17 + 25 comparison using related FAPs
23 + 27 [146]
24 + 25 comparison using related FAPs
24 + 26 comparison using related FAPs
24 + 27 comparison using related FAPs
20 + 22 comparison using related FAPs
20 + 23 comparison using related FAPs
227
APPENDIX B. MAPPING FROM FAPS TO AUS
228
Appendix C
229
APPENDIX C. EMOTION VALUES IN THE ACTIVATION-EVALUATION SPACE
230
Table C.1 Continue
Emotion Activation Evaluation
Jealous 6.1 3.4
Joyful 5.4 6.1
Loathful 3.5 2.9
Lonely 3.9 3.3
Meek 3.0 4.3
Nervous 5.9 3.1
Obedient 3.1 4.7
Obliging 2.7 3.0
Outraged 4.3 3.2
Panicky 5.4 3.6
Patient 3.3 3.8
Pensive 3.2 5.0
Pleased 5.3 5.1
Possessive 4.7 2.8
Proud 4.3 5.7
Puzzled 2.6 3.8
Quarrelsome 4.6 2.6
Rebellious 5.2 4.0
Rejected 5.0 2.9
Remorseful 3.1 2.2
Resentful 5.1 3.0
Sad 3.8 2.4
Sarcastic 4.8 2.7
Satisfied 4.1 4.9
Scornful 5.4 4.9
Self-controlled 4.4 5.5
Serene 4.3 4.9
Sociable 4.8 5.3
Sorrowful 4.5 3.1
Stubborn 4.9 3.1
Submissive 3.4 3.1
Surprised 6.5 5.2
Suspicious 6.5 5.2
Sympathetic 3.6 3.2
Terrified 6.3 3.4
Trusting 3.4 5.2
Unaffectionate 3.6 2.1
Unfriendly 4.3 1.6
Wondering 3.3 5.2
Worried 3.9 2.9
231
APPENDIX C. EMOTION VALUES IN THE ACTIVATION-EVALUATION SPACE
232
Appendix D
This chapter contains the articles that were presented and published in different con-
ferences and journals, with the objective to present the partial results we were achieving
along the realization of this thesis.
233